Search

Image processing is fundamental to observational astronomy workflows. Astronomers acquire imaging data, and process the imagery to extract useful information. This thesis introduces two new image processing algorithms. The first, PyTorchDIA, is a GPU-accelerated approach to Difference Image Analysis (DIA). The approach is fast, without sacrificing modelling flexibility. It makes use of the Pythonic, PyTorch machine learning framework to accelerate convolution computations on the GPU, and compute gradients of user-specified objective functions with automatic differentiation methods to fit DIA models quickly and accurately. The second algorithm, The Thresher, was designed as a new tool to extracting information from Lucky Imaging (LI) data sets. We adopt a modelling approach which optimises a justifiable, physically motivated likelihood function to return the best estimate of the observed astronomical scene. It does this using all available data, and the more data the model is fit to, the better the signal-to-noise and resolution of the scene estimate. This fundamentally differs from conventional shift-and-add procedures, which typically reject the vast majority of the acquired LI data, as in these approaches, the signal-to-noise of the final coadd is inversely related to its resolution. With an eye to accessibility, integration into workflows and open science, the code for these two algorithms has been open sourced. Lastly, we show how Python image processing applications can be used to realise time-critical, demanding computational challenges in a chapter outlining the results of a novel pilot study to detect the occultations of background stars by small, outer solar system objects with high frame-rate sCMOS cameras.

A circadian clock is present in some capacity in almost all forms of life, and is useful for a wide array of traits, but crucially allows organisms to predict future conditions and adapt their behaviour to synchronise with, and thrive under, their dynamic environment. Accordingly, plant environmental stress responses are gated in a circadian manner, including that for pathogenic defence and immunity. Comparatively less work has been carried out on the plant pathogens, but there are increasing reports of pathogens capable of rhythmically altering their behaviour and virulence-related traits. Magnaporthe oryzae, the fungal pathogen responsible for the destruction of enough rice to feed at least 60 M people annually, has been shown to possess some circadian clock components, and based on bioinformatic analyses, likely contains all the core, accessory, and circadian-associated genes. M. oryzae displays a conidial banding pattern, reminiscent of the model clock species, N. crassa, and (after sufficient entrainment) this pattern can continue to occur under free running conditions for a number of days, with a period of approximately 24 h. This rhythm is also presented on a range of nutrient-rich and poor media, suggesting a nutritionally-compensated circadian rhythm in M. oryzae. This onset of conidial banding is partially determined by the presence of secreted metabolites, the sensation of which is facilitated by the circadian clock, predominantly via WC2. The entraining light conditions that M. oryzae is exposed to can significantly alter its vegetative growth, conidiation and conidial development, and even pathogenicity. Further, inoculation timing (dawn or dusk) plays a role in both the virulence of M. oryzae, and in the susceptibility of the plant host, seemingly in a species-by-species manner, where rice is most susceptible at dawn, and barley most susceptible at dusk. For M. oryzae, pre-inoculation entrainment to darkness predominantly favours dawn inoculations, and those exposed to prolonged periods of light prefer dusk inoculation. Upon mutation of the core clock genes, WC2 and FRQ, vegetative growth, conidiation and conidial development, photoadaptation, and pathogenicity were all significantly altered compared to the wild type, suggesting an important role of the clock in the general fitness of M. oryzae. This work discusses how entraining light cycles and the circadian clock impacts the growth, development, conidiation, virulence, and ultimate severity in the economically important rice blast disease.

Current methods of Meibomian gland imaging present analysis at a single point in time, with the assumption that gland structure changes very slowly. Despite the importance of meibography in Meibomian gland dysfunction diagnosis, management, and treatment there is still poor evidence on whether gland structure or the presence of functional acini is demonstrated in meibography images. Meibography is a fully objective technique, however the image interpretation remains subjective and unclear. This thesis reports on work to enhance our understanding of Meibomian gland physiology via the development of improved imaging systems, image analysis software and clinical studies evaluating change in gland appearance over time. The first clinical study reports on diurnal and monthly variations of visible Meibomian gland structure. The findings revealed some variation over time when glands are being analysed individually. To further investigate changes in Meibomian gland appearance, the second clinical study explored how Meibomian glands change after therapeutic expression. It was shown that Meibomian glands get dimmer and shorter following therapeutic expression but recover within 24 hours meaning that they are capable of 'regeneration' in a relatively short period of time. Development of custom imaging system and image analysis software helped capture and analyse images in a reliable way. The results presented in this thesis add to our understanding of individual Meibomian gland function in a young and healthy population. This information is essential to fully understand the complicated nature of holocrine glands by providing insights into whether meibography captures gland structure or the presence of functional acini. Furthermore, the reported deep learning-based algorithm can be utilised in Meibomian gland health assessment by facilitating clinicians in Meibomian gland dysfunction diagnosis, management and treatment.

Radiotherapy is a cancer therapy in which ionising radiation is used to target cancerous tissue within the human body. The therapeutic aim is to irradiate cancerous tissue and create DNA strand breaks such that the cancerous tissue is unable to self-repair and subsequently dies. The primary therapeutic consideration is to minimise healthy tissue irradiation such that the probability of secondary cancers is minimised. External Beam Radiotherapy (EBRT) is a form of radiotherapy which uses a linear particle accelerator (LINAC) to generate an X-ray treatment beam to target cancerous tissue deep within the human body. Advanced forms of X-ray radiotherapy such as Intensity Modulated Radiotherapy (IMRT) utilises an advanced collimation device, known as a multileaf collimator (MLC) which is composed of thin abutting tungsten slats referred to as leaves to shape the X-ray treatment beam such that it conforms to the tumour geometry. The precision of IMRT and other advanced EBRT's is largely dependent upon an MLC's ability to precisely shape the X-ray treatment beam. It is standard practise to calibrate MLC leaves to a tolerance of ±1 mm every 3 months. To maintain total dose errors below 2%, the verification tolerance should be set to a higher standard of ±0.3 mm. Pre-treatment verification, which is recommended for each treatment to minimise errors is time consuming, and is thus seldom performed for each individual patient, but instead for an acceptable sample of patients and/or treatment-deliveries. Incorporating a high-precision, real-time treatment monitoring device would allow MLC errors to be detected instantaneously and eliminate the need for pre-treatment verification as each treatment would be verified in real time. This would increase patient safety and treatment quality. Monolithic Active-Pixel Sensors (MAPS) are thin silicon pixel sensors with a small pixel size. They therefore have low attenuation and can be used for high resolution measurements. They can also be made to be radiation hard and can have a high speed readout. This makes MAPS an excellent candidate for real-time monitoring in radiotherapy. In this thesis a MAPS-based device is presented and shown to have excellent performance as a real-time, upstream treatment verification device. It is shown that the position of MLC leaves can be reconstructed with resolutions ranging between 62±6µm and 86±16µm depending on the leaf configuration using 0.15 sec of treatment data. An upstream full-scale large area MAPS device is tested and shown to have a clinically insignificant attenuation such that it does not change the delivered treatment when used as a monitoring device. Although the gap between the abutted sensors makes the leaf position resolution worse, it is still well within clinical tolerance. The device was shown to be radiation hard under proton irradiation. It can operate after radiation-induced proton damages of up to 20 kGy at 20°C, and up to 50 kGy and potentially beyond when cooled to -2.13±0.461°C or below. This is equivalent to ~ 2 years of clinical deployment. The work in this thesis demonstrates the device known as the demonstrator is a clinically deployable solution as a real-time, upstream treatment verification device.

In the UK, around 100,000 people have a stroke each year, equating to one stroke every five minutes. Prompt treatment is required to give a good patient outcome. To help decide what the treatment options are, most suspected stroke patients in the UK have a brain scan performed on admission to hospital. Each scan is examined by a radiologist, and a report describing the images is written. An automatic artificial intelligence (AI) solution for brain scan analysis could support the radiologist to report scans quickly and accurately, in turn helping with the speed and accuracy of treatment. However, training AI models for medical image analysis requires large amounts of expertly annotated data which are time-consuming and expensive to obtain. Large unlabelled datasets such as those for stroke patients are currently difficult to use directly for training. An emerging solution is to leverage radiology reports as a source of expert information in order to construct imaging training annotations. In this thesis, a system for extracting suitable labels from radiology reports of suspected stroke patients is proposed and implemented in collaboration with NHS Greater Glasgow and Clyde. A state-of-the-art deep learning natural language processing (NLP) model is first trained to extract the labels from radiology reports automatically by combining per-label attention with a novel data augmentation approach that uses templates and knowledge bases. This NLP model is used to automatically create training annotations from an uncurated dataset of over 27,000 reports, larger than is realistically feasible using only manual annotations. Finally, an image analysis model is trained to detect haemorrhagic regions in the brain. Using automatically extracted annotations improves performance over manual annotations by 20%. Being able to use these data without the need for time-consuming and expensive annotation efforts to be carried out by already stretched healthcare professionals provides a promising avenue for transformative medical image analysis research.

RGB-D cameras are a type of imaging device that includes a standard colour camera and a depth sensor. The depth sensor records depth images, where each pixel measures the distance of the target from the camera. Using data from the RGB-D camera, we can recon- struct a three-dimensional scene. Since the introduction of the Microsoft Kinect, multiple affordable RGB-D cameras have reached the market and are even integrated into laptops and mobile phones. In this thesis, healthcare applications of RGB-D cameras are investigated, focussing on two areas. The first is Structured Light Plethysmography (SLP), a method for measuring lung function. The non-contact nature of SLP is ideal for respiratory measurement, as it avoids potential contamination without requiring single-use equipment. The current SLP device is large and expensive, while the data it produces is low-resolution and does not adapt to the patient. A new SLP processing pipeline based on an RGB-D camera is proposed, which will produce automatically aligned, high-resolution data in a portable device. Algorithm development to enhance the data processing capabilities of this new pipeline is presented. A method for interpreting data from the high-resolution grid, via a control point-based shape modelling technique, is used to show differences in the respiratory pat- tern before and after exercise. A system is described for compensating for the motion of the subject, so the resulting data only shows motion due to breathing. This method is demonstrated by compensating for the pedalling motion of a subject on an exercise bike. The second application area is the Neonatal Intensive Care Unit (NICU), which cares for critically ill new-born babies. A proof-of-concept study is designed and conducted to investigate the potential use of an RGB-D camera in the NICU. In this thesis, the collec- tion of 24 hours of RGB-D video footage is presented, and it is shown how the infant's respiratory rate can be measured from the depth data.

In cancer therapy, tumour hypoxia is an important predictor of disease progression, treatment outcomes, and metastatic potential. To date, methods for reducing tumour hypoxia have delivered limited success owing to either a lack of proven efficacy, unwanted side effects, or logistical limitations. Therefore, there remains an urgent requirement for an effective hypoxia reduction mechanism that is safe, cost effective, biocompatible, and easy to use. To address this, oxygen-filled micro and nanobubbles have been investigated as a method of oxygen delivery and have shown promising results in improving therapeutic response. To develop and optimise oxygen carriers and other hypoxia reduction strategies, however, consistent methods for measuring changes in tissue oxygenation need to be established. The ideal method should be noninvasive and quantitative, allowing tumour oxygen measurements to be obtained before, during, and after treatment. This thesis examines whether clinically available MRI techniques can be used to detect the change in tissue oxygenation caused by the nano-carriers. A combination of phantom and preclinical experiments, mathematical modelling, and the analysis of data from two clinical trials are used. First, original results from phantom and preliminary in vivo MRI experiments are reported, which demonstrate the feasibility of detecting the effects of oxygen carriers. Second, three separate mathematical models are developed and presented in order of increasing complexity 2 first addressing simple bodily fluids, then blood, and finally, tissues 2 each of which are useful for areas of MRI research beyond the focus of this thesis. Third, original research results from recently completed clinical studies of oxygen-enhanced MRI are analysed and used to assess the viability of using these MRI measurements to monitor oxygen delivery from nanobubbles in a clinical MRI setting. Lastly, the broader potential applications of the conclusions and tools presented in this thesis are discussed.

The main objective of this thesis is to develop a better understanding of materials' interfaces in spatial and temporal domains over dynamical industrial processes via 2D and 3D visualization and quantitative acoustic measurements using Ultrasound Computed Tomography (USCT). USCT is an imaging method that permits the determination of the spatial distribution of materials based on their acoustic properties. Tomographic techniques are usually based on a number of sensors that are located in a meaningful order around the region-of-interest (ROI). In the particular tomographic technique, piezoelectric transducers that emit signals of a standard frequency have been used. The main advantage of this imaging modality is the low-cost development but also its non-destructive and non-invasive behaviour to the process. The control of industrial processes is becoming crucial as the research and engineering community aims at fully automated solutions. Thus, monitoring and control of tank reactors play a significant role in the overall developments. In general, this thesis focuses on improving the performance of static/dynamical USCT visualization and quantitative measurements for industrial processes. The developed studies aim to investigate in depth the USCT developments and the application in industrial reactors processes and dynamical complex processes involving high mixing. This work aims especially, to investigate USCT functionality in the batch crystallization process, where there is a great need for real-time measurement and integration of imaging and control systems towards the process automation. Detailed applications and further work are suggested. Although USCT is a new imaging technique and lots of issues are still preserved to be solved, it is believed that such research would have contributions to the future development of USCT research. This thesis includes a number of novel approaches that constitute contributions to the USCT imaging field. At first the quantitative sound-speed imaging was developed, providing good correlation with concentration in concentrated solutions. Then the same concept transmission imaging tested in many crystallization scenarios establishing a thorough investigation of USCT in batch crystallization. In continuation, the developed transmission imaging extended in 3D and finally in 4D imaging using a spatiotemporal Total Variation algorithm. Lastly, a new triple-modality imaging is introduced as a fast and accurate method for industrial imaging proving to be more efficient than the single methods.

Over the past decade, the overlap between machine learning and image processing has grown so considerably that the two fields have become inseparable. Some of the most important problems affecting science and society fall within these overlapping fields, including the development of self-driving cars and the automated interpretation of medical images. Innovation in machine learning and image processing is boosted by two pillars: extraordinarily complex models, and extremely large datasets. Training complex models on large datasets is challenging, and research developing efficient algorithms to do so has grown in parallel with the rise in machine learning and image processing. Stochastic optimisation algorithms have long been the methods of choice for training machine learning models. For reasons outlined in this thesis, they are natural choices for solving classical problems in machine learning, and their efficiency generally scales with the size of training datasets much better than other algorithms. Stochastic gradient methods have been studied extensively in regulated scenarios: stochastic gradient estimators are generally required to be unbiased, and the optimisation objectives often must be strongly convex with simple non-smooth regularisers (if any non-smoothness is permitted). As the applications of machine learning expand, especially within image processing, their associated optimisation problems become more complex, and researchers can no longer rely on existing theoretical guarantees established under highly controlled scenarios. New optimisation algorithms must be developed to efficiently solve problems at the intersection of machine learning and imaging. This thesis seeks to improve our understanding of stochastic optimisation algorithms and expand these algorithms into new domains, with an emphasis on problems from image processing. We introduce new optimisation algorithms and develop theoretical frameworks to study these methods without assuming unbiased estimators, strong convexity, or other restrictive assumptions that most existing algorithms require. With this flexibility, we in- troduce several algorithms that exhibit state-of-the-art efficiency on problems in machine learning and image processing. Many of the algorithms introduced in this thesis achieve theoretically optimal convergence rates on their respective problem classes. By solving large-scale problems more efficiently, we hope the algorithms in this thesis allow researchers to pursue even more complex problems and explore even larger datasets. With the theoretical frameworks in this thesis, we hope to provide the foundation for a better understanding of the behaviour of optimisation algorithms in machine learning in image processing.

Rotational moulding is a manufacturing process that produces huge hollow plastic parts compared to injection or blow moulding. In a typical rotomoulding cycle, a polymer powder is loaded into a metal mould to be bi-axially rotated and heated. Once a molten plastic layer entirely covers the mould, the tool is moved to a cooling chamber where the temperature decreases until the part is solidified to be de-moulded. Rotational moulding is well known in the tank manufacturing industry as it is one of the most extended techniques for manufacturing large plastic tanks. Also, rotomoulding can produce articles such as shipping containers, recycling bins, boats or kayaks. Despite its advantages, the method's full potential is not exploited due to the limited number of materials available for moulding and their mechanical properties. During the classical rotomoulding cycle, specifically during heating, the powder particles undertake a process in which they soften and merge to form a homogeneous layer. This phenomenon is driven by surface tension and is known as sintering, which constitutes the fundamental mechanism by which parts are formed in rotomoulding. The complexity of sintering makes adding and combining different materials to gain part strength incredibly challenging. Combining two or more materials can create compounds of different properties, known as composites, that produce a result structure that offers more desirable properties than the separate sources. However, no method has made standard the incorporation of reinforcements to produce composite structures using rotational moulding. The present work aims to develop effective means to reinforce rotationally moulded parts to enhance strength and mechanical performance. The work is divided into three main blocks. In block one, the work starts with determining the optimum material composition and structure. An experimental study identifies the critical parameters of the composite part's optimum material composition and structure. The study also aids in developing a hybrid manufacturing process by defining the addition method, processing parameters and moulding technique. An experimental study focuses on how polymer particles soften, bond, and densify with reinforcing fibres. Adhesion measurements were based on the angle formed between the liquefied resins and the reinforcing fibre bundle surface. It was discovered that by modifying heating conditions, resin viscosity, and fibre chemical surface, the adhesion of the polymer to the fibre could be wettable enhanced. In block two, the work continues developing a new in-mould visual imaging system, named RotoCam™, capable of obtaining images of the polymer during the process. The high temperatures, bi-axial rotation, and enclosed mould complicate studying the polymer melt from the inside of the mould. An investigation into the system's main characteristics is undertaken, identifying the key features it requires and assessing the image quality and level of magnification. The layouts included developing three versions, achieving the optimum image quality and resolution to acquire measurable images from the process in the last design. The work finishes by applying the optimum material composition, structure, and hybrid manufacturing process in a full-scale set of trials in block three. During the trials, RotoCam™ acquired images to assess fibre-polymer adhesion, validating the material and the process. The trial set comprised mouldings using the resins, fibres and processing parameters established in block one to determine the relation between fibre adhesion with composite fibre content. Finally, conclusions and recommendations for future work are presented in the last chapter to produce the most optimum composite structure in rotational moulding.

Electricity networks are critical national infrastructure throughout the world, delivering vital energy services and supporting interdependent assets such as transport and hospitals. The network corridors are inspected and refurnished regularly to remain useful. Due to the high number of components and the geographical spread of electricity networks, operating costs can be very high. Inspection parameters include vegetation encroachment, sagging lines, tower paintwork defects and numerous components on the towers. There are over fifty inspection parameters that are critical along a network segment. The state-of-the-art inspection process involves aerial surveys. Images are acquired and analysed manually, which adds to the high cost of aerial surveys. In addition to being costly, this process can suffer from inter-observer and intra-observer variability. The condition-based risk management model is a popular network asset management model within the industry. The model allows for individual components rating and then enables a collective economic impact analysis for the medium and long term. The assessment model has two main stages that inform if a tower requires intervention. An aim of the routine analysis stage is to rate towers, select "at-risk" candidates for detailed investigation and refurbishment. The effectiveness of the assessment model would depend on how quick and accurate the routine assessment is able to highlight areas of need considering the limited resources during the inspection window. This thesis focuses on automating tasks within the routine analysis stage involving image analysis. Specifically, this thesis identified towers that are at-risk of different failure modes using deep learning. The first step in the proposed pipeline involves the identification of tower types as suspension and tension. We found that towers could be categorised automatically by focusing on the configurations at tower cross-arms. In addition to tower type detection, identifying images of cross-arm, body and foot would serve as precursor for effective extraction of specific inspection parameters. Components such as anti-climbing devices are found around tower body and not higher up the cross-arm or peak of the tower. Hence, classifying tower images to reflect the region of interest would provide a filter for object detection. Tower conditions are often associated with the failure modes of components instances they support. This thesis demonstrated an automated detection of insulators and U-bolts as exemplar inspection parameters. Our method classified at-risk towers based on the detected instances but without explicitly labelling the instances. Instead, learning was supervised using only the condition labels of towers in their entirety. This enables us to use a real-world industry dataset without the requirement of fine-grained data annotation of thousands of individual components. While existing studies classified component instances on a tower, we classified the tower as a whole and show that tower labels are adequate for the task. Automated detection and analysis of U-bolts have not been previously reported. Furthermore, the thesis presents the identification of paintwork deterioration for image-based tower management. We argue that the classification of tower parts may be costly. While identifying towers with immediate need for intervention, our approach to tower paintwork classification could be used as an early warning system for assets approaching end-of-life. The utility of each sub-system has been demonstrated using a real-world industry dataset of over 7k towers and 300k images that are representative of asset failure modes and inspection scenarios.

The life-threatening acute coronary syndromes (ACSs) are caused mainly by the rupture of the coronary atherosclerotic plaques. Angiography-defined luminal stenosis is the only validated criterion for clinical decision making. However, post-mortem studies showed that only 14% of myocardial infarctions (MIs) were caused by plaques with >70% stenosis. Plaque vulnerability, characterised by morphological and compositional features, is more important than the luminal stenosis measurements. Mini-invasive virtual-histology intravascular ultrasound (VH-IVUS), optical coherence tomography (OCT), computed tomography angiography (CTA), magnetic resonance imaging (MRI) and positron emission tomography (PET) have been developed to characterise lesion morphology, composition, function and metabolism. However, the overall predictive power of an imaging-defined high risk feature for future ischaemic events is insufficient for clinical decision making. New biomarkers are therefore needed. Coronary atherosclerosis is a chronic inflammatory disease, and the plaque is subject to mechanical loading due to dynamic blood pressure and flow, as well as bending due to heart motion. The mechanical loading estimation and the inflammation quantification have the potential in generating new biomarkers for vulnerability assessment. An accurate mechanical analysis depends on many factors, such as material properties, loading conditions and modelling strategies. The influence of material properties and loading conditions has been widely studied, whilst modelling strategies are currently less investigated. A PET/MR system provides simultaneous anatomical and metabolic imaging. However, accurate attenuation correction is essential for quantitative PET reconstruction in the thorax, and the application of PET/MR in coronary imaging is limited due to the challenges of obtaining good quality images of the coronary vessels. The work presented in this thesis first investigates the influence of computation strategies and bending on the calculation of mechanical parameters, in particular, plaque structure stress (PSS). The hypothesis is that the 3D structure only finite element analysis (FEA) can accurately predict the PSS with good time efficiency, and bending is an important factor in coronary biomechanical analysis. Mechanical parameters calculated from the 3D fully coupled fluidstructure interaction (FSI) model with the patient-specific bending were used as references. 3D structure-only FEA showed a good agreement with the gold standard model and a shorter solution time (Chapter 3). The comparison of mechanical features with and without patientspecific coronary bending showed an increase in PSS due to the bending and revealed the importance of bending in coronary FEA (Chapter 4). For the imaging part, a hybrid bias correction method was proposed for the thoracic region using zero echo time (ZTE) images, which were converted to pseudo-CT images of the lung region. These images were integrated with the standard MR attenuation map to improve the accuracy of attenuation correction (AC) in PET reconstruction (Chapter 5). We hypothesise that non-Cartesian MR sequences could improve the coronary image quality in a PET/MR system, and the incorporation of ZTE based pseudo-CT in the lung can improve the PET attenuation correction. The imaging part of this work involved the development of 3D nonCartesian (spiral and radial) trajectory MR sequences for the PET/MR system (Chapter 6). Overall, the 3D structure only FEA with patient-specific bending can accurately estimate the coronary PSS with a reasonable computation time. The ZTE based pseudo-CT attenuation correction in the lung region can improve the accuracy of thoracic PET reconstruction. The application of non-Cartesian sequences for coronary imaging requires further development in the PET/MR system.

Medical Imaging, which allows for the non-invasive assessment of biological tissues, is a rapidly growing health care service that has evolved from a diagnostic tool to a platform for personalized precision medicine. Computerized Tomography (CT), which is a commonly obtained imaging study worldwide, utilizes X-ray radiation to differentiate tissues based on density differences. More complex imaging studies are obtained based on the clinical question to supplement the CT. Image overutilization, which is the acquisition of medical images that have minimal impact on patient care and increasing global inequities due to the cost/availability of medical imaging services are significant problems that need to be addressed. One possible solution would be to maximize the amount of clinically relevant information extracted from routine CT images using machine learning approaches. Recent developments in machine and deep learning have provided a powerful set of tools for the automated and complex analysis of medical images. Leveraging these techniques, it is possible to develop algorithms capable of learning from human annotations or paired images. These methods can also be applied to large datasets to minimise human input and to enable the rapid scale up of analyses. In this thesis, I hypothesize that Non-Contrast CT images contain higher-order information to differentiate tissue anatomy or pathology without the need of intravenous contrast agents or radioactive tracers. I further hypothesized that such higher-order information enables stratification of disease progression without the need of additional imaging studies. I focus on two pathologies, abdominal aortic aneurysms (AAA) and Head and Neck Squamous Cell Carcinoma (HNSCC) as demonstration of feasibility. The methods described in this thesis can be applied to other pathologies and will be poised to disrupt clinical pathways in the future.

In recent years, a lot of interest has been generated in structural health inspection automation. Although a number of computer-vision based crack detection techniques have been designed but further research is needed. Firstly, they are crack focused so there is a requirement to research for Pits. Secondly, these techniques have not been used in practical visual inspection as they are often developed under near-ideal conditions, not in an on-site environment, which contain more complexities such as variable shapes, glare, and loss of focus due to curved surfaces. Thirdly, there is a need to detect pits and pit-to-cracks of microscopic level for early assessments. This research presents an automated inspection technique that is able to perform detection and measurements of on-surface micro-flaws caused by corrosion fatigue, including classification of pits and cracks by using artificial intelligence (AI) technology. It is an industry-driven application that can be used for both rail axles and pipelines. It has been used for two purposes. First purpose, is to serve as part of a tool to enable early detection of cracks and assess the remaining service life of a corroded axle. The measurement of the damage depends on the detection of small microscopic cracks (around 0.1-0.3mm long) within the corroded area. This serves as an input value into the remaining-life software in the RAAI project. The results have been validated by the Polimi data showing that the implemented method is able to provide a prompt outcome including highlighting location, measuring and counting specific features, using on-site data. The second purpose, is to create a tool that can provide assistance to a corrosion assessment operator. The outcome includes highlighting, measuring and counting specific features plus it is able to classify between pits and cracks, using on-site data. It also implements industrial standard, considering the closeness of the pits, their size and density. Thus, the tool reduces the skill level requirement of an operator, as the algorithm sets a standard for the desired defects to be counted in quantitative measures. The applicability of the proposed method has been evaluated on images taken from the field. The evaluation results confirm the high adoptability of the proposed method for defect inspection in an on-site environment. Firstly, new databases have been designed and created for the project that includes data handling, data gathering and data labelling. Database of such nature doesn't exist per existing research for the specific research problem. The data has been gathered by multiple sources that includes three site visits as well as data by performing bending testing at TWI, which expands the depth of the database. Pixel-wise labelled database of 165,888,000 data inputs, consisting of 115 microscopic pit images and 20 crack images has been created for this research task. Secondly, advance unsupervised image segmentation based detection, localisation, measurement and assessment is built that produces better results than the state-of-the-art algorithm, for this specific industry-driven problem. Watershed and Morphological-based (shape) algorithms are implemented for analysing and assessment. This includes tasks such as finding shapes, detecting edges, removing noise, object detection, counting objects, segmentation, filtering and region analysis. It is able to show quantitative results such as number of flaws detected, along with the flaw's length, area, eccentricity, and their location shown on the image. They were then tested at different sites such as data from Ireland, Italy and UK; and validated with real data such as number of flaws, their average length and the longest flaw. Detailed indicative measures were applied to test and validate the system's performance such as 95.2% accuracy, 55% precision, f1 score 56%, probabilistic rand Index (PRI) 91.7%, CV is 42.8%, VOI as low as 41.08%, and Global consistency error (GCE) as low as 2.6%. Thirdly, novel supervised machine learning methods are proposed. Especially, deep learning model is implemented as an image-wise classification model, pre-trained with AlexNet. The resultant outcome displays the class of the image to which it belongs. If the image has a pit, then the Pit model is able to pick it with 91.4% accuracy, and if the image has a crack, it is picked by the Crack model with a high accuracy of 98%. Another deep learning model implemented is a pixel-wise segmentation, based on a combination of two state-of the art models, UNet with VGG16. It shows performance with global accuracy of 93%. With a validation accuracy of 95% on validation training dataset and testing mean accuracy of 91%, mean IoU is 63.71% with a weighted IoU being 90.4%. Lastly and most importantly, a Defect Detection System (DDS) has been designed, implemented, tested and verified in a real-industrial application. It is a practical solution with a ready-to-use and resource-efficient design. The setup requires a laptop, portable microscope and an automated scanner. The laptop is attached to the microscope which capture flaws with up to 0.08mm size length sensitivity and saves data in images/video format. The microscope is mounted on the scanner, that auto-rotates the camera circumferential as well as in axial axes, along the structure component being inspected. It is based on the combination of both supervised and unsupervised learning methods by merging their strengths. Detects, measures and localises by unsupervised image segmentation, so it doesn't need to perform lengthy pixel-wise labelling; and classifies the flaws by using deep learning so it doesn't need to hard-code complex computational values. It is simple yet efficient.

Expressing ideas in our minds which are inevitably visual into words had been a necessity. Lack of tools to express oneself creatively using visual communication had been a bottleneck. With the advent of AI and generative models starting to be applied to graphics based tasks, this elusive elysian fields seems to be a step closer. This thesis tries to tackle the challenge using the 3 pronged approach of GANs, VAEs and Continuous Normalizing Flows. We first propose MAD-GAN, an intuitive generalization to the Generative Adversarial Networks (GANs) and its conditional variants to address the well known problem of mode collapse. First, MAD-GAN is a multi-agent GAN architecture incorporating multiple generators and one discriminator. Second, to enforce that different generators capture diverse high probability modes, the discriminator of MAD-GAN is designed such that along with finding the real and fake samples, it is also required to identify the generator that generated the given fake sample. Intuitively, to succeed in this task, the discriminator must learn to push different generators towards different identifiable modes. We perform extensive experiments on synthetic and real datasets and compare MAD-GAN with different variants of GAN. We show high quality diverse sample generations for challenging tasks such as image to image translation and face generation. In addition, we also show that MAD-GAN is able to disentangle different modalities when trained using highly challenging diverse-class dataset (For example the dataset with images of forests, icebergs, and bedrooms). In the end, we show its efficacy on the unsupervised feature representation task. We introduce a similarity based competing objective (MAD-GAN-Sim) which encourages different generators to generate diverse samples based on a user defined similarity metric. We show its performance on the image-to-image translation, and also show its effectiveness on the unsupervised feature representation task. We then propose an interactive GAN-based sketch-to-image translation method that helps novice users create images of simple objects. As the user starts to draw a sketch of a desired object type, the network interactively recommends plausible completions, and shows a corresponding synthesized image to the user. This enables a feedback loop, where the user can edit their sketch based on the network's recommendations, visualizing both the completed shape and final rendered image while they draw. In order to use a single trained model across a wide array of object classes, we introduce a gating-based approach for class conditioning, which allows us to generate distinct classes without feature mixing, from a single generator network. In Appendix C we propose a semi supervision based strategy for deep generative modelling for human body analysis. Deep generative modelling for human body analysis is an emerging problem with many interesting applications. However, the latent space learned by such approaches is typically not interpretable, resulting in less flexibility. In this work, we present deep generative models for human body analysis in which the body pose and the visual appearance are disentangled. Such a disentanglement allows independent manipulation of pose and appearance, and hence enables applications such as pose-transfer without specific training for such a task. Our proposed models, the Conditional-DGPose and the Semi-DGPose, have different characteristics. In the first, body pose labels are taken as conditioners, from a fully-supervised training set. In the second, our structured semi-supervised approach allows for pose estimation to be performed by the model itself and relaxes the need for labelled data. Therefore, the Semi-DGPose aims for the joint understanding and generation of people in images. It is not only capable of mapping images to interpretable latent representations but also able to map these representations back to the image space. We compare our models with relevant baselines, the ClothNet-Body and the Pose Guided Person Generation networks, demonstrating their merits on the Human3.6M, ChictopiaPlus and DeepFashion benchmarks. Finally, Training Neural Ordinary Differential Equations (ODEs) is often computationally expensive. Indeed, computing the forward pass of such models involves solving an ODE which can become arbitrarily complex during training. Recent works have shown that regularizing the dynamics of the ODE can partially alleviate this. In this paper we propose a new regularization technique: randomly sampling the end time of the ODE during training. The proposed regularization is simple to implement, has negligible overhead and is effective across a wide variety of tasks. Further, the technique is orthogonal to several other methods proposed to regularize the dynamics of ODEs and as such can be used in conjunction with them. We show through experiments on normalizing flows, time series models and image recognition that the proposed regularization can significantly decrease training time and even improve performance over baseline models.

When submerged in agarose gels, growing Escherichia coli (E. coli) form lenticular colonies. In this thesis I first present a study on the origin of this morphology. Using single plane illumination microscopy, I record the full 3D morphology of growing E. coli colonies for the first time. By adding fiducial markers to the gel, I measured the displacement of the gel surrounding the growing colonies. These measurements revealed that as E. coli undergo binary fission, growing from a single bacterium into a colony of ∼ 10⁶ cells, they fracture the gel around them. However, this fracture process cannot be described by simple linear elastic fracture mechanics. Instead, fractures opened up by expanding colonies undergo a transition in propagation mode, changing from one where the gel is fractured over the entire crack surface, to one where the gel is fractured in a single plane. To my knowledge, this is the first time this transition has been measured, showing the potential of using bacteria to explore small scale fracture mechanics. The observed transition in fracture mode is consistent with cavitation theory, suggesting that at small scales the pressure required for colony growth is independent of the gel's fracture energy. This independence may mean that bacteria can grow in far tougher materials than would previously have been expected. In a second study I develop an apparatus that can inject oil bubbles into agarose at the same length and time scales as bacterial colonies grow, effectively creating a physical simulation of colony growth. With this apparatus I show that there is little difference between fractures propagated by a Newtonian fluid and those propagated by a colony of E. coli. In fact, what morphological differences could be observed can be plausibly explained by the difference in interfacial energy between the agarose-E. coli and agarose-oil interface. This means that in the future hydraulic fracture theories may be applied wholesale to predictions of bacterial infiltration into materials. In a final study, the elastic and fracture properties of agarose were measured using a custom apparatus to perform Rivlin and Thomas's pure shear test. The fracture energy was found to be in reasonable agreement with the theory of Lake and Thomas, being of order 0.4 J m−² for 2 % agarose. The fracture energy was found to be independent of the fracture rate below cross head speeds of 0.01 mm s−¹, meaning that the viscous dissipation in the gel surrounding the colonies is small and that the measurements made in this study are applicable to predictions of the slowly growing colony morphology.

In recent years, 3D displays have been recognized as the ultimate dream of immersive display technology and there have been a great development immersive 3D technology including AR/VR and auto-stereoscopic 3D displays. Holoscopic 3D (H3D) system is one of the autostereoscopic 3D which is a true 3D imaging principle which mimics fly's eye technique to capture and replay using a micro lens array which is an array of perspective lens of the same specification. LED wall display has shown a fast growth where LED digital displays are widely used in both in/outdoor for advertisement and entertainment. Ultra-big LED display monitor is an ideal hardware device to provide remarkable 3D viewing experience and fit numbers of viewers to perceive 3D effects at same time. However, compare with existing 3D technologies which successfully applied on LCD display monitor, LED display still suffers from resolution when applied pixel mapping method which uses number of 2D pixels to construct a 3D pixel. In this PhD research, an innovative 3D pixel mapping was explored and designed to enhance 3D viewing experience in horizontal direction of LED 3D Wall-size display. In particular, an innovative Holoscopic 3D imaging principle is used to design and prototype LED 3D Wall display of resolution enhancement. Compare with the classic 3D display method, this enhanced display method of LED display improved horizontal resolution double times without losing any viewpoints. The outcome research is promising as a good depth and motion parallax for medium to long distance viewing are achieved. In addition to the aforementioned, to improve the quality of rendered 3D images of LED display in omnidirectional directions, a distributed pixel mapping algorithm was designed to reduce the lens pitch three times to gain smoother motion parallax of rendered 3D images compare with traditional pixel mapping method in omnidirectional direction. Unfortunately, due to lack of high-resolution LED display monitor, this distributed pixel mapping method was eventually tested and evaluated on LCD display with 4K resolution.

This study aimed to develop a novel end-to-end plant diagnosis model for the analysis of plant health conditions in near real-time to optimize the rate of production on farmlands for an intensive, yet environmentally safe farming production to preserve the natural environment. First, field research was conducted to determine the extent of the problems faced by farmers in agricultural production. This allowed us to refine the research statement and the level of technology involved in the production processes. The advantages of unmanned aerial systems were exploited in the continuous monitoring of farm plantations to develop automated and accurate measures of farm conditions. To this end, this thesis applies the Precision Agricultural technology as a data based management system that takes into account spatial variations by using the Global Positioning System, Geographical Information System, remote sensing, yield monitors, mapping, and guidance system for variable rate applications. An unmanned aerial vehicle embedded with an optic and radiometric sensor was used to obtain high spectral resolution images of plantation status during normal production/growth cycle. Then, an ensemble of classifiers with Convolution Neural Networks (CNN) was used as off the shelf feature extractor to train images to develop an end-to-end feature detection and multiclass classification system for plant overall health's conditions. Whereby previous works have concentrated on using CNN as off the shelf feature extractor and model training to detect only plant diseases from plants. To date, no research has yet been carried out to develop an end-to-end model for the overall plant diagnosis system. Previous studies focused on the detection of diseases at any given time, making it difficult to implement comprehensive real-time PA systems. Applying the pretrained model to the new images showed that the model can accurately predict any plant condition with an average of 97% accuracy.

Cardiovascular magnetic resonance (CMR) imaging is a powerful tool for research and clinical applications. To extract useful clinical information from the acquired CMR images, time-consuming and laborious manual delineation of cardiovascular structures is currently required. Despite promising overall performance across medical imaging applications, the current state-of-the-art automated image segmentation methods still fail in some cases, potentially jeopardising the reliability of clinical diagnosis. Thus, it is important to develop not only automation of image segmentation but also quality control of segmentation, to empower efficient and reliable CMR image data analysis. To address both segmentation and quality control problems, I have developed a novel quality control-driven (QCD) framework in this thesis. Extending upon deep ensemble learning, the framework utilises multiple convolutional neural network-based models to generate segmentation candidates, the agreement of which is exploited via additional regression models to predict segmentation quality measured by Dice similarity coefficient (DSC). The DSC prediction not only provides a quality estimate but also enables a novel approach to select a final, most optimal segmentation on-the-fly from multiple candidates, improving segmentation robustness. Following the DSC prediction, a segmentation quality classification scheme is implemented to alert human operators only when manual intervention is recommended, intended for more efficient allocation of time and labour resources for large-scale image processing pipelines. Through both quantitative and qualitative evaluation, the QCD framework has demonstrated excellent performance in both segmentation and quality control. More importantly, the framework has been successfully applied across CMR imaging techniques, anatomical structures, and large-scale datasets acquired at different sites, with high adaptability and generalisability. The QCD framework could pave the way towards large-scale automated imaging data analysis pipelines, with both efficiency and reliability, in real-world clinical applications.

Cardiac magnetic resonance (CMR) is the current gold standard imaging technique to assess myocardium anatomy and regional and global function of the heart. The accuracy and robustness of CMR imaging has been validated using manual segmentation. However, the quality of manually traced contours is critical for the assessment. Manual segmentation is tedious and time consuming and suffers from inter- and intra-observer variability. Therefore, much research work has been focused on automating the segmentation of CMR images. The rapid developments and applications of deep learning algorithms have dramatically improved the accuracy of automated segmentation methods and open the possibilities for alternative approaches. In this thesis, we first investigate and show the importance of the segmentation method for calculating cardiac anatomy and function parameters using deep neural networks. We then develop and validate our segmentation neural network using both human and mouse CMR images for both short-axis (SAX) slices and long-axis (LAX) slices in order to utilise all available information for more meaningful and accurate evaluation. We then develop our three-dimensional (3D) neural network for reconstructing ventricular surfaces using sparse contours extracted from both SAX and LAX two-dimensional (2D) slices, with or without spatial consistency between them. Finally, we integrate our segmentation and reconstruction networks into a fully automated pipeline for the assessment of mouse heart using CMR images and achieved promising results by validating against our database with both normal and infarct hearts coming from two imaging centres.

Neural circuit models explain an animal's behavior as evoked activity of different circuit elements of its nervous system. Synaptic wiring diagrams mapped from structural imaging of nervous systems guide modeling of neural circuits on the basis of connectivity. However, connectivity alone may not sufficiently constrain the set of plausible circuit hypotheses for empirical study. Combining structural imaging of synaptic connectivity with functional information from activity imaging can further constrain these hypotheses to create unequivocal neural circuit models. This thesis develops computational methods and tools to cross-reference structural and activity imaging of explant larval Drosophila central nervous systems at cellular resolution. Augmenting synaptic wiring diagrams with activity maps via these methods relates circuit structure and function at the neuronal level on a per-behavior basis. Neuronal activity of larval central nervous systems expressing pan-neuronal calcium indicators is imaged in a light sheet microscope, which are then structurally imaged with high throughput electron microscopy. Methods and tools are provided for the assembly of these image volumes, spatial registration between imaging modalities, automated detection of relevant tissue and cellular structures in each, extraction of activity time series, and morphological identification of neurons in structural imaging using reference wiring diagrams mapped from other larvae. Using these methods, existing wiring diagrams mapped from a reference first instar larva were identified with neurons in a larva augmented with activity information for a neural circuit involved in peristaltic motor control. This demonstrates the feasibility of the contributed methods to associate the wiring diagrams of arbitrary circuits of interest with activity timeseries across multiple individuals, behaviors, and behavioral bouts. To demonstrate capability to augment wiring diagrams with information besides activity, these methods are also applied to multiple larvae each expressing specific neurotransmitter labels rather than calcium indicators in the light sheet microscopy. This work scaffolds future modeling of circuits underlying behavior that can only be mechanistically understood in the context of multi-modal observation of synaptic connectivity, functional activity and molecular markers. The methods developed also enable comparative connectomics between multiple individuals, which is necessary to study inter-individual variability in circuits and to observe experimental intervention in the development, structure, and function of neural circuits.

Magnetic resonance imaging (MRI) is a notoriously slow imaging method compared with other imaging modalities, for example computed tomography. It can, however, produce contrasts and attain biophysical information that is unattainable by other methods and is thus a valuable clinical tool. In recent years, image acquisition times have been reduced through image reconstruction methods that require less data than traditional methods. In this thesis, these kinds of methods are used and extended upon for improved acquisition speed of an MRI method that is particularly slow due to requiring multiple encoded acquisitions to produce a single composite image: arterial spin labelling (ASL). In conventional ASL, two encodings are used; one with "labelled" blood and one without "labelling", such that subtracting one from the other gives an image of just the blood signal. Angiography can be used to visualise blood flow through the arteries, and perfusion imaging to assess oxygen and nutrient supply to the tissue. Some advanced ASL methods require even more encodings that can be decoded to reveal more information about the cerebral haemodynamics. An example of such an advanced ASL method is vessel-encoded ASL, which allows for generation of separate images of blood originating from different arteries. In this thesis, modern MRI sampling and reconstruction methods are optimised for vessel-encoded ASL and other ASL variants, with the aim of bringing these modalities towards clinically feasible scan times, which would eventually allow for more information-rich assessments of the cerebrovasculature and the perfusion state of the brain in, for example, patients suffering from stroke, dementia, and arteriovenous malformations. By careful joint consideration of the multi-dimensional data, the data acquisition, and the reconstruction, very high acceleration factors can be achieved. This is demonstrated, first for vessel-encoded ASL angiography in 2D and 3D, then in similar advanced ASL methods (time-encoded ASL angiography, and combined angiography and perfusion imaging). A new radial sampling scheme is also presented and assessed on ASL angiographic data, that could have impact on imaging methods beyond ASL, in particular other dynamic MRI modalities.

The Arabic language originally made use of ancient calligraphy, which is found in historical documents and the Holy Quran. This calligraphy shows a different representation of Arabic texts using a more cursive style and a mixture of complex constructed word forms. These types of writing styles in Arabic texts give a degree of difficulty in segmenting the letters and reading the text. Since they originated long ago, they have mostly been reflected in Islamic culture and the use of quotations from the Holy Quran. This art form is still used today for various purposes in Arabic representation and Islamic calligraphy. The challenges of this type of text motivate the search for a way to simplify the reading and digitisation processes. To the best of our knowledge, this is the first attempt to investigate the recognition of Arabic calligraphy images and the reading of text drawn in such images. Due to the lack of resources in the calligraphy domain, different datasets were developed for this research. An Arabic calligraphy image dataset was collected, from which calligraphy letter image datasets were generated. Finally, a calligraphy quotations corpus was manually annotated based on the Arabic calligraphy image dataset. All of these datasets were used for training, testing and support in the different phases that were applied to achieve the primary goal of reading calligraphy. A new approach to the recognition of Arabic calligraphy was developed to manipulate a scanned image and extract a list of probable quotations. It consists of comparing two detection methods, namely maximally stable extremal regions (MSER) and sliding window (SW), to obtain the identity of intersecting letters from the image. The letters detected had their features extracted through a comparison between the histogram of oriented gradient (HOG) features and a bag of speeded-up robust features (SURF) used in training two different recognition models, support vector machines (SVM) and the convolutional neural network (CNN). In the investigation into which of these models and image feature descriptors most accurately fit the calligraphy letters, the results from the recognition process were placed in a bag of letters (BOL) feature. This feature was used to search the corpus according to two different methodologies to produce the list of probable quotations. The first method compares the target BOL with the corpus index BOL for each element, while the second method generates a list of related words from the BOL and then searches the corpus for any quotations that contain two or more of these words. The results from reading 388 calligraphy images showed that the MSER method outperforms the SW method in detecting letters. Moreover, BOL matching and searching the corpus predicts more accurate lists of quotations than the word generation process. The best methodology is based on a combination of the SVM recognition model and HOG feature extraction, correctly predicting more than 74% of the top ten quotations using the BOL matching process.

Developing FPGA-based applications is typically a slow and multi-skilled task. Tools to support application development have gradually become more high level. This thesis describes an approach which aims to raise further the level at which an application developer works while developing FPGA-based implementations of image and video processing applications. The starting concept is that of a soft coprocessor (SCP). A set of fundamental soft coprocessors is presented, aimed at implementing some of the key abstractions of Image Algebra. Each soft coprocessor processes an image or video stream, and can be parameterised by its specific functionality. To support interactive experimentation, the developer can edit the functionality and parameters without requiring re-synthesis in most cases. Our soft coprocessors are designed to enable easy chaining of SCPs, as defined by a Dataflow Graph (DFG). This reduces considerably the amount of data transfer between SCPs and the host processor. Although hardware optimisation is not a primary concern of this thesis, further contributions facilitate the development of more efficient and more specific versions of the more generic soft coprocessors, based on the concept of reusable coprocessor skeletons. To facilitate the creation of more algorithm-specific coprocessors which cannot be accommodated within the skeleton framework, a layered library of macros is developed, which enables a developer to create new coprocessors without requiring extensive hardware design knowledge. The thesis illustrates these novel concepts by presenting soft coprocessors for a range of common image processing algorithms, including a novel fast implementation of K-means Clustering for grey level images, GLCM feature extraction, and Histogram of oriented Gradients (HoG). Finally, a prototype implementation of an integrated development environment, called SCoPeS, is presented. This enables a user to express an application as a text-based DFG of soft coprocessors, and automatically obtain and run an FPGA implementation. The ability to modify an application, even during execution, without requiring re-synthesis, is an important benefit given the experimental nature of image and video processing application development.

A computer aided diagnosis system (CAD) is developed to fully characterize and classify mass to benign and malignancy and to predict BIRAD (Breast Imaging Reporting and Data system) scores using mammographic image data. The CAD includes a preprocessing step to de-noise mammograms. This is followed by an active counter segmentation to deforms an initial curve, annotated by a radiologist, to separate and define the boundary of a mass from background. A feature extraction scheme wasthen used to fully characterize a mass by extraction of the most relevant features that have a large impact on the outcome of a patient biopsy. For this thirty-five medical and mathematical features based on intensity, shape and texture associated to the mass were extracted. Several feature selection schemes were then applied to select the most dominant features for use in next step, classification. Finally, a hierarchical classification schemes were applied on those subset of features to firstly classify mass to benign (mass with BIRAD score 2) and malignant mass (mass with BIRAD score over 4), and secondly to sub classify mass with BIRAD score over 4 to three classes (BIRAD with score 4a,4b,4c). Accuracy of segmentation performance were evaluated by calculating the degree of overlapping between the active counter segmentation and the manual segmentation, and the result was 98.5%. Also reproducibility of active counter 3 using different manual initialization of algorithm by three radiologists were assessed and result was 99.5%. Classification performance was evaluated using one hundred sixty masses (80 masses with BRAD score 2 and 80 mass with BIRAD score over4). The best result for classification of data to benign and malignance was found using a combination of sequential forward floating feature (SFFS) selection and a boosted tree hybrid classifier with Ada boost ensemble method, decision tree learner type and 100 learners' regression tree classifier, achieving 100% sensitivity and specificity in hold out method, 99.4% in cross validation method and 98.62 % average accuracy in cross validation method. For further sub classification of eighty malignance data with BIRAD score of over 4 (30 mass with BIRAD score 4a,30 masses with BIRAD score 4b and 20 masses with BIRAD score 4c), the best result achieved using the boosted tree with ensemble method bag, decision tree learner type with 200 learners Classification, achieving 100% sensitivity and specificity in hold out method, 98.8% accuracy and 98.41% average accuracy for ten times run in cross validation method. Beside those 160 masses (BIRAD score 2 and over 4) 13 masses with BIRAD score 3 were gathered. Which means patient is recommended to be tested in another medical imaging technique and also is recommended to do follow-up in six months. The CAD system was trained with mass with BIRAD score 2 and over 4 also 4 it was further tested using 13 masses with a BIRAD score of 3 and the CAD results are shown to agree with the radiologist's classification after confirming in six months follow up. The present results demonstrate high sensitivity and specificity of the proposed CAD system compared to prior research. The present research is therefore intended to make contributions to the field by proposing a novel CAD system, consists of series of well-selected image processing algorithms, to firstly classify mass to benign or malignancy, secondly sub classify BIRAD 4 to three groups and finally to interpret BIRAD 3 to BIRAD 2 without a need of follow up study.

Image understanding has long been a goal for computer vision. It has proved to be an exceptionally difficult task due to the large amounts of variability that are inherent to objects in a scene. Recent advances in supervised learning methods, particularly convolutional neural networks (CNNs), have pushed forth the frontier of what we have been able to train computers to do. Despite their successes, the mechanics of how these networks are able to recognize objects are little understood, and the networks themselves are often very difficult and time-consuming to train. It is very important that we improve our current approaches in every way possible. A CNN is built from connecting many learned convolutional layers in series. These convolutional layers are fairly crude in terms of signal processing - they are arbitrary taps of a finite impulse response filter, learned through stochastic gradient descent from random initial conditions. We believe that if we reformulate the problem, we may achieve many insights and benefits in training CNNs. Noting that modern CNNs are mostly viewed from and analyzed in the spatial domain, this thesis aims to view the convolutional layers in the frequency domain (viewing things in the frequency domain has proved useful in the past for denoising, filter design, compression and many other tasks). In particular, we use complex wavelets (rather than the Fourier transform or the discrete wavelet transform) as basis functions to reformulate image understanding with deep networks. In this thesis, we explore the most popular and well-developed form of using complex wavelets in deep learning, the ScatterNet from Stephane Mallat. We explore its current limitations by building a DeScatterNet and found that while it has many nice properties, it may not be sensitive to the most appropriate shapes for understanding natural images. We then develop a locally invariant convolutional layer, a combination of a complex wavelet transform, a modulus operation, and a learned mixing. To do this, we derive backpropagation equations and allow gradients to flow back through the (previously fixed) ScatterNet front end. Connecting several such locally invariant layers allows us to build learnable ScatterNet, a more flexible and general form of the ScatterNet (while still maintaining its desired properties). We show that the learnable ScatterNet can provide significant improvements over the regular ScatterNet when being used as a front end for a learning system. Additionally, we show that the locally invariant convolutional layer can directly replace convolutional layers in a deep CNN (and not just at the front-end). The locally invariant convolutional layers naturally downsample the input (because of the complex modulus) while increasing the channel dimension (because of the multiple wavelet orientations used). This is an operation that often happens in a CNN by a combination of a pooling and convolutional layer. It was at these locations in a CNN where the learnable ScatterNet performed best, implying it may be useful as learnable pooling layer. Finally, we develop a system to learn complex weights that act directly on the wavelet coefficients of signals, in place of a convolutional layer. We call this layer the wavelet gain layer and show it can be used alongside convolutional layers. The network designer may then choose to learn in the pixel or wavelet domains. This layer shows a lot of promise and affords more control over what regions of the frequency space we want our layer to learn from. Our experiments show that it can improve on learning in the pixel domain for early layers of a CNN.

In recent years there has been an increased interest on flexible fluid-structure interaction problems with applications to flow control (reduction of drag and lift fluctuations) and energy harvesting. Particularly, studies have suggested that a hairy coating (poroelastic coating) may help reduce drag and lift fluctuations of a bluff body (cylinder) by around 15% and 40%, respectively. However, these studies have focused on two-dimensional setups, therefore real effects such as three-dimensional vortex shedding in the wake of a cylinder have not been considered. For instance, preliminary experiments carried out in a wind tunnel revealed that the motions of a single filament in the wake of a cylinder are complex due to the influence of the cylinder wake flow and the outer crossflow impinging on the hanging filament (sagged due to gravity). For this reason, this work was set to study experimentally the fundamental behaviour of filaments alone, hanging from a vertical support tube (i.e. not attached to a bluff body). This simple configuration is ideal to analyse the fundamental dynamics of flexible filaments in flow and provide insights for future investigations of hairy coatings. Noteworthy, the filaments hanging in crossflow were free to move in three dimensions, in contrast to the previously existent studies which come from two-dimensional studies, and thus provides unprecedented data valuable for validating fluid-structure interaction simulation codes. At low wind speeds the filaments bent and remained in static equilibrium, similar to the reconfiguration of plants. Beyond this condition, at a certain wind speed, the filaments started to vibrate and in certain cases entered into large-amplitude three-dimensional flutter motions which became more complicated as the wind speed was further increased. Through the use of stereoscopic non-contact high-speed imaging, hotwire anemometry, smoke visualizations and the recourse to linear and nonlinear time-series analysis techniques, the full range of filament behaviours were studied in detail. In particular, the results from this research provided unprecedented data and empirical correlations for the filament static reconfiguration and fluid loading at previously unexplored conditions. Also, the fluid mechanisms responsible for the onset of filament motion were investigated. Additionally, the vortex shedding from reconfigured filaments was for the first time experimentally studied and characterized. This work also provided the first documentation of the three-dimensional flutter motions of filaments, and the effects of turbulence intensity and filament attachment angle on the filaments flapping motions dynamics. Finally, the experimental methodologies (data acquisition, image processing and time-series analyses of motion) developed during this research were also applied for studying other fluid-structure interactions problems: the flow-induced vibration of cantilever rods in axial flow for nuclear reactor applications, and the dynamics of inverted flags for energy harvesting applications.

This thesis is concerned with methods for Bayesian inference and their applications in astrophysics. We principally discuss two related themes: advances in nested sampling (Chapters 3 to 5), and Bayesian sparse reconstruction of signals from noisy data (Chapters 6 and 7). Nested sampling is a popular method for Bayesian computation which is widely used in astrophysics. Following the introduction and background material in Chapters 1 and 2, Chapter 3 analyses the sampling errors in nested sampling parameter estimation and presents a method for estimating them numerically for a single nested sampling calculation. Chapter 4 introduces diagnostic tests for detecting when software has not performed the nested sampling algorithm accurately, for example due to missing a mode in a multimodal posterior. The uncertainty estimates and diagnostics in Chapters 3 and 4 are implemented in the $\texttt{nestcheck}$ software package, and both chapters describe an astronomical application of the techniques introduced. Chapter 5 describes dynamic nested sampling: a generalisation of the nested sampling algorithm which can produce large improvements in computational efficiency compared to standard nested sampling. We have implemented dynamic nested sampling in the $\texttt{dyPolyChord}$ and $\texttt{perfectns}$ software packages. Chapter 6 presents a principled Bayesian framework for signal reconstruction, in which the signal is modelled by basis functions whose number (and form, if required) is determined by the data themselves. This approach is based on a Bayesian interpretation of conventional sparse reconstruction and regularisation techniques, in which sparsity is imposed through priors via Bayesian model selection. We demonstrate our method for noisy 1- and 2-dimensional signals, including examples of processing astronomical images. The numerical implementation uses dynamic nested sampling, and uncertainties are calculated using the methods introduced in Chapters 3 and 4. Chapter 7 applies our Bayesian sparse reconstruction framework to artificial neural networks, where it allows the optimum network architecture to be determined by treating the number of nodes and hidden layers as parameters. We conclude by suggesting possible areas of future research in Chapter 8.

This thesis is concerned with accurate imaging from radio interferometry data and with subsequent analysis so as to determine source positions and fluxes in the radio sky. The thesis makes proposals and implementations of new algorithms, which substantially improve the accuracy of image products and the results of source extraction. These improvements in accuracy promise to assist scientific research into astronomical objects and phenomena in radio astronomy. The thesis contains six chapters, beginning with an overview of the imaging process in radio interferometry in Chapter 1. Chapter 2 focuses on improving the accuracy of source extraction, by utilising the Bayesian methodology. The proposed Bayesian method has been implemented in a software package called 'BaSC' which uses the Markov Chain Monte Carlo (MCMC) technique. By design, it works with intermediate radio interferometry image products, such as dirty images, rather than with reconstructed images. BaSC achieves greater precision in source location and better resolving power than mainstream source extraction software such as SExtractor, which works with reconstructed images. This finding confirms that reconstructed images are not a true representation of the radio sky, whereas dirty images already contain full information about the observations. This piece of work has been accepted by Monthly Notices of the Royal Astronomical Society (Hague et al. 2018). Chapter 2 is based on this paper, but has been rewritten and expanded. Based on this conclusion, Chapter 3 seeks to optimise the gridding process so as to make accurate dirty images. Since the Fast Fourier transform (FFT) produces dirty images with a much lower computational cost than the Direct Fourier transform (DFT), a new gridding function is needed which minimises the difference between DFT and FFT dirty images. The 'Least-misfit' gridding function is proposed, so as to minimise the image misfit between the DFT and FFT dirty images, and this is implemented and tested. Given an identical support width, it outperforms the main-stream spheroidal function in the image misfit by a factor of at least 100, it also suppresses aliasing in the image plane better. Aliasing is essentially a part of the image misfit, so there is no need to consider it separately. The least-misfit gridding function, with a support width of 7 and an image cropping rate of 0.5, is recommended for application to both the gridding and degridding processes, and makes it realistic to achieve the limit of single precision arithmetic in the image misfit and visibility misfit. With the new gridding function in place, Chapter 4 proposes two novel wide-field imaging algorithms: improved W-Stacking and N-Faceting. The improved W-Stacking method uses a three-dimensional gridding, rather than two-dimensional gridding as in the original W-Stacking method. This renders possible the calculation and application of the correcting function on the n (third) dimension. This improvement greatly increases the accuracy of the FFT dirty image on the celestial sphere, relative to the DFT dirty image. The image misfit is as small as 10⁻⁸ when using the proposed least-misfit gridding function with a support width of 8, and it further reaches the double precision limit when the support width is increased to 14. For comparison, the image misfit levels achieved by the W-Projection algorithm in CASA and the original W-Stacking algorithm in WSCLEAN are 10⁻³, several orders of magnitude worse. In addition, since the number of w-planes required by the improved W-Stacking method is reduced compared to the original method, cutting a significant amount of FFT computational cost. As for the original W-Stacking method, if less w-planes than needed are used, the dirty images and reconstructed images produced will underestimate the fluxes of sources that are located far from the phase centre. The N-Faceting method involves imaging of multiple n-planes, followed by a three-dimensional deconvolution process, where a position-independent beam is used. Chapter 5 applies the improved W-Stacking method to two real sets of observational data, comprising one GMRT dataset and one VLA dataset. The dirty images on the celestial sphere and the reconstructed images are shown. The improved W-Stacking method successfully removes non-coplanar effects. For the observation with a larger range of w, improved W-Stacking method is recommended to generate a more accurate image with lower computational cost compared to the original W-Stacking method. Finally, Chapter 6 sets out conclusions drawn from the present work.

Microfluidic droplet systems have shown great promise in the study of biological systems. In this thesis I explore the application of a microfluidic droplet system to the study of small bacterial populations and their growth response to antibiotics. This thesis comprises of three primary results sections. The first section details the development and fabrication of a custom microfluidic system. This microfluidic system is designed to study the growth dynamics of small Escherichia coli (E. coli ) bacterial communities. The second section presents an image processing and analysis workflow that is designed to be used in conjunction with the microfluidic device to extract quantitative growth data of confined bacterial communities with single cell precision. The third section details a study that uses the microfluidic system to measure the growth dynamics of E. coli communities encapsulated within microfluidic droplets whilst uninhibited and under the effect of the antibiotic streptomycin. In the first section, I present the development, design and fabrication of a two-part microfluidic device that is later used to study the growth dynamics of E. coli bacteria. The first part of the microfluidic device consists of droplet generating microfluidic geometry that is designed to encapsulate small communities of E. coli bacteria (typically 1-5 cells in size) within microfluidic droplets around 30μm in diameter. The second part of the microfluidic device is composed of a micro fluidic reservoir that is designed to store microfluidic droplets. By imaging the droplets in the storage reservoir with a combination of brightfield and fluorescence microscopy, the growth dynamics of the encapsulated bacterial communities can be described. In the second section of the thesis, I present the development and operating principles of my image processing and analysis workflow. This workflow automatically extracts quantitative bacterial growth data from a gridded array of brightfield and fluorescence microscopy images taken across the microfluidic device during experimentation. The analysis algorithm is capable of processing upwards of 200 fields of view across 80 time frames. This makes it possible to detect, track and measure the size of upwards of 1000 encapsulated bacterial communities with single cell precision. Droplet boundaries are detected using brightfield microscopy images and each droplet is tracked from one time step to the next using a modified particle tracking algorithm. Bacterial community sizes are measured by counting individual bacteria in thresholded fluorescence microscopy images. In the final section of this thesis, a study of E. coli bacteria was conducted using the microfluidic device. Small communities of bacteria, typically 1-5 cells in size, were confined in microfluidic droplets and observed for 8 hours. The bacteria were grown in the absence of antibiotic, as well as in the presence of the antibiotic streptomycin at various concentrations. A number of metrics of the growth dynamics are extracted from the study and compared with traditional bulk growth techniques. It is shown that in the absence of streptomycin, the final droplet population size distributions feature extended tails due to large, fast growing bacterial populations. The tails of these distributions are reduced in length in the presence of streptomycin at a concentration below the Minimum Inhibitory Concentration (MIC). A positive correlation is found between initial and final bacterial population size, however this correlation is weak, indicating a highly stochastic growth process. It is also shown that at moderate streptomycin concentrations both around and above the MIC, some bacterial populations are inhibited in a non-lytic manner, whilst other populations underwent lysis. It is also shown that at moderate antibiotic concentrations around the MIC value, there is a bias against bacterial lysis in droplets with large populations. Finally, at low antibiotic concentrations below the MIC, it is shown that some bacterial populations are inhibited but not lysed, whereas some populations are not inhibited. The microfluidic device and image analysis workflow described in this thesis is a novel experimental system with widespread applicability. The system is able to partition an aqueous solution into a large number of small, compartmentalised liquid volumes and analyse these volumes using both brightfield and fluorescence time-lapse microscopy. This could be used to study the population dynamics of a variety of bacterial strains, and could also be used to explore more general questions relating to stochastic cell biology where encapsulation is experimentally beneficial.

The breakup of large rocket or satellite structures during destructive re-entry is not well understood and depends heavily on numerical simulations to predict the entry-trajectories. Experimental studies in high-speed wind tunnels provide validation data for such simulations with non-intrusive optical methods extracting quantified field data and true free-flight capabilities of object motion in free-flight. Optical schlieren have become the de facto method to visualise the flow structures around objects in high-speed flow. It affords high spatiotemporal resolution from precision optics and both still and high-speed cameras. It is not possible to simultaneously resolve multiple-strength density gradients in a single schlieren image, which requires either multiple costly tunnel runs to tune knife edge orientation and cut-off which cannot precisely duplicate transient flow features. The background-oriented schlieren (BOS) method allows a single image to be interrogated for multiple knife edge orientations and different dynamic range values. Two sets of experiments were performed in this study; steady-state and free-flight. Steady-state tests were performed to evaluate the spatial resolution, sensitivity and dynamic range of schlieren and (BOS) applied to hypersonic flows with multiple density gradients. Metrics were devised for optimisation of the physical BOS arrangement to minimise the shock and model blur and maximise the shock displacement. A processing study developed metrics to achieve minimal processing blur and highest signal to noise ratio. The BOS results have an overall lower spatial resolution and sharpness than the schlieren images, but had lower shock blur than schlieren for the optimised arrangement used. A synthetic knife edge and nonlinear colour were applied to the BOS studies to resolve an increased number of weaker flow features in a single image than multiple orientations of optical schlieren knife edge and cut off settings. These methods provided increased spatial resolution of the density gradients in a BOS image than a standard linear colourmap for the full range of displacement vectors. For thefree-flight testing both BOS and optical schlieren captured the six distinct Edney shock interaction types. Application of nonlinear colourmaps and synthetic knife edge orientations improved resolution of weaker shear-layer and boundary layer separation flow features behind the sphere than captured with the schlieren. Sphere free-flight drag coefficient was reconstructed using moving linear fit to derive acceleration. Uncertainty with this approach was less than 3%. A lower drag coefficient was seen for the sphere in free-flight but this was comparable to studies with high-sphere surface roughness. Spheres experiencing an Edney shock interaction produced an increased drag coefficient with a maximum value for the Edney IV case. Expanding the applicability range of non-intrusive optical field measurement techniques for high-speed flows is vital to support experimental testing and validation of numerical research results. Significant developmental work was performed on the BOS technique to evaluate its performance, sensitivity, blur and expand its dynamic rage to support future experimental tests in conjunction with the established schlieren methods.

Computer vision has seen great advances recently due to artificial neural networks (ANNs) with multiple hidden layers; in particular, convolutional neural networks (CNNs) have shown great flexibility in the tasks they are able to perform. These networks are usually trained in a supervised manner using the back-propagation algorithm. Spiking neural networks (SNNs) are the third generation of neuron simulation and have been proven to have better mathematical properties than previous generations. This approach to simulating biological neurons uses pulses to communicate the state of the units. The event-driven nature of SNNs imposes challenges for computer vision tasks. For example, neuromorphic vision sensors generate data when changes in their field of view are perceived; in contrast, standard cameras send full images. Furthermore, SNN training methodologies are not as mature as previous generations' procedures; critically, biologically-plausible supervised learning algorithms are scarce. In this thesis we explore event-based computer vision using spiking neurons as computation units; the main goal is to build a fully-spiking visual pipeline. Firstly, we develop methods to transform standard images into spike representations. We then process the generated spike trains to extract salient features using biological principles. Finally, supervised learning algorithms for SNNs are developed and applied in a computer vision context.

Digital forensics practitioners can be tasked with analysing digital data, in all its forms, for legal proceedings. In law enforcement, this largely involves searching for contraband media, such as illegal images and videos, on a wide array of electronic devices. Unfortunately, law enforcement agencies are often under-resourced and under-staffed, while the volume of digital evidence, and number of investigations, continues to rise each year, contributing to large investigative backlogs. A primary bottleneck in forensic processing can be the speed at which data is acquired from a disk or network, which can be mitigated with data reduction techniques. The data reduction approach in this thesis uses reduced representations for individual images which can be used in lieu of cryptographic hashes for the automatic detection of illegal media. These approaches can facilitate reduced forensic processing times, faster investigation turnaround, and a reduction in the investigative backlog. Reduced file representations are achieved in two ways. The first approach is to generate signatures from partial files, where highly discriminative features are analysed, while reading as little of the file as possible. Such signatures can be generated using either header features of a particular file format, or by reading logical data blocks. This works best when reading from the end of the file. These sub-file signatures are particularly effective on solid state drives and networked drives, reducing processing times by up to 70× compared to full file cryptographic hashing. Overall the thesis shows that these signatures are highly discriminative, or unique, at the million image scale, and are thus suitable for the forensic context. This approach is effectively a starting point for developing forensics techniques which leverage the performance characteristics of non-mechanical media, allowing for evidence on flash based devices to be processed more efficiently. The second approach makes use of thumbnails, particularly those stored in the Windows thumbnail cache database. A method was developed which allows for image previews for an entire computer to be parsed in less than 20 seconds using cryptographic hashes, effecting rapid triage. The use of perceptual hashing allows for variations between operating systems to be accounted for, while also allowing for small image modifications to be captured in an analysis. This approach is not computationally expensive but has the potential to flag illegal media in seconds, rather than an hour in traditional triage, making a good starting point for investigations of illegal media.

Title: The subnuclear localisation of Notch responsive genes. Candidate Name: Matthew Jones Notch signalling is a highly conserved cell-cell communication pathway with critical roles in metazoan development and mutations in Notch pathway components are implicated in many types of cancer. Notch is an excellent and well-studied model of biological signalling and gene regulation, with a single intracellular messenger, one receptor and two ligands in Drosophila. However, despite the limited number of chemical players involved, a striking number of different outcomes arise. Molecular studies have shown that Notch activates different targets in different cell types and it is well known that Notch is important for maintaining a stem cell fate in some situations and driving differentiation in others. Thus some of the factors affecting the regulation of Notch target genes are yet to be discovered. Previous studies in various organisms have found that the location of a gene within the nucleus is important for its regulation and genome reorganisation can occur following gene activation or during development. Therefore this project aimed to label individual Notch responsive loci and determine their subnuclear localisation. In order to tag loci of interest a CRISPR/Cas9 genome-editing method was established that enabled the insertion of locus tags at Notch targets, namely the well-characterized Enhancer of split locus and also dpn and Hey, two transcription factors involved in neural cell fate decisions. The ParB/Int system is a recently developed locus tagging system and is not well characterised in Drosophila. It has a number of advantages over the traditional LacO/LacI-GFP locus tagging system as it does not rely on binding site repeats for signal amplification and can label two loci simultaneously in different colours. This thesis characterised the ParB/Int system in the Drosophila salivary gland and larval L3 neuroblast. Using 3D image segmentation hundreds of nuclei were reconstructed and a volume based normalisation method was applied to determine the subnuclear localisation of several Notch targets with and without genetic manipulations of the Notch pathway.

The subject of this thesis is a class of fast inpainting methods (image or video) based on the idea of filling the inpainting domain in successive shells from its boundary inwards. Image pixels (or video voxels) are filled by assigning them a color equal to a weighted average of either their already filled neighbors (the ``direct'' form of the method) or those neighbors plus additional neighbors within the current shell (the ``semi-implicit'' form). In the direct form, pixels (voxels) in the current shell may be filled independently, but in the semi-implicit form they are filled simultaneously by solving a linear system. We focus in this thesis mainly on the image inpainting case, where the literature contains several methods corresponding to the {\em direct} form of the method - the semi-implicit form is introduced for the first time here. These methods effectively differ only in the order in which pixels (voxels) are filled, the weights used for averaging, and the neighborhood that is averaged over. All of them are very fast, but at the same time all of them leave undesirable artifacts such as ``kinking'' (bending) or blurring of extrapolated isophotes. This thesis has two main goals. First, we introduce new algorithms within this class, which are aimed at reducing or eliminating these artifacts, and also target a specific application - the 3D conversion of images and film. The first part of this thesis will be concerned with introducing 3D conversion as well as Guidefill, a method in the above class adapted to the inpainting problems arising in 3D conversion. However, the second and more significant goal of this thesis is to study these algorithms as a class. In particular, we develop a mathematical theory aimed at understanding the origins of artifacts mentioned. Through this, we seek is to understand which artifacts can be eliminated (and how), and which artifacts are inevitable (and why). Most of the thesis is occupied with this second goal. Our theory is based on two separate limits - the first is a {\em continuum} limit, in which the pixel width →0, and in which the algorithm converges to a partial differential equation. The second is an asymptotic limit in which h is very small but non-zero. This latter limit, which is based on a connection to random walks, relates the inpainted solution to a type of discrete convolution. The former is useful for studying kinking artifacts, while the latter is useful for studying blur. Although all the theoretical work has been done in the context of image inpainting, experimental evidence is presented suggesting a simple generalization to video. Finally, in the last part of the thesis we explore shell-based video inpainting. In particular, we introduce spacetime transport, which is a natural generalization of the ideas of Guidefill and its predecessor, coherence transport, to three dimensions (two spatial dimensions plus one time dimension). Spacetime transport is shown to have much in common with shell-based image inpainting methods. In particular, kinking and blur artifacts persist, and the former of these may be alleviated in exactly the same way as in two dimensions. At the same time, spacetime transport is shown to be related to optical flow based video inpainting. In particular, a connection is derived between spacetime transport and a generalized Lucas-Kanade optical flow that does not distinguish between time and space.

Radiotherapy planning involves the calculation of dose deposition throughout the patient, based upon quantitative electron density images from computed tomography (CT) scans taken before treatment. Cone beam CT (CBCT), consisting of a point source and flat panel detector, is often built onto radiotherapy delivery machines and used during a treatment session to ensure alignment of the patient to the plan. If the plan could be recalculated throughout the course of treatment, then margins of uncertainty and toxicity to healthy tissues could be reduced. CBCT reconstructions are normally too poor to be used as the basis of planning however, due to their insufficient sampling, beam hardening and high level of scatter. In this work, we investigate reconstruction techniques to enable dose calculation from CBCT. Firstly, we develop an iterative method for directly inferring electron density from the raw X-ray measurements, which is robust to both low doses and polyenergetic artefacts from hard bone and metallic implants. Secondly, we supplement this with a fast integrated scatter model, also able to take into account the polyenergetic nature of the diagnostic X-ray source. Finally, we demonstrate the ability to provide accurate dose calculation using our methodology from numerical and physical experiments. Not only does this unlock the capability to perform CBCT radiotherapy planning, offering more targeted and less toxic treatment, but the developed techniques are also applicable and beneficial for many other CT applications.

Particles exist in great abundance in nature, such as in sands and clays, and they also constitute 75% of the materials used in industry (e.g., mineral ores, formulated pharmaceuticals, dyes, detergent powders). When a load is applied to a bulk assembly of soil particles, the response of a geomaterial at the bulk (macro) scale, originates from the changes that take place at the particle scale. If particle breakage occurs, the shape and size of the particles comprising the bulk are changed; this induces changes in the contact network through which applied loads are transmitted. As a result, changes at the micro-scale can significantly affect the mechanical behaviour of a geomaterial at a macro-scale. It is therefore unsurprising that the mechanisms leading to particle breakage are a subject of intense research interest in several fields, including geomechanics. In this thesis, particle breakage of two dry granular materials is studied, both experimentally and numerically. The response of the materials is investigated under different stress paths and in all the tests grain breakage occurs. High resolution x-ray computed micro-tomography (XCT) is used to obtain 3D images of entire specimens during high confinement triaxial compression tests and strain controlled oedometric compression tests. The acquired images are processed and measurements are made of the temporal and spatial evolution of breakage, local variations of porosity, volumetric and shear strain and grading. The evolution and spatial distribution of quantified breakage including the resulting particle size distribution for the whole specimen and for specific areas, are presented and further related to the localised shear and volumetric strains that developed in the specimens. In addition, the discrete element method (DEM) was used to provide further micro-mechanical insight of the underlying mechanisms leading to particle breakage. Classical DEM simulations, using a Hertz-Mindlin contact model and non-breakable spheres, was first deployed to study the initiation and likelihood of particle breakage under oedometric compression. Moreover, a bonded DEM model was used to create clumps that represent each particle and simulate breakage of particles under single particle compression. The DEM model parameters were obtained from results of single particle compression test and the models were validated against the quantitative 3D information of the micro-scale, acquired from the XCT analysis.

Structure-from-motion (SfM) photogrammetry has become ubiquitous in the geosciences, owing to its low-cost and ease of use for generating 3D data. Ideas around data collection, quality and processing need to be revisited to ensure that the technology is being harnessed correctly. One area which is new in this current image acquisition boom is the range of sensors and systems being used to collect image data. This raises crucial questions in the geoscience community which are addressed in this contribution. This is split into three parts. Firstly, image quality is investigated to establish whether a stable association between it and the quality of photogrammetric products can be uncovered to allow simpler and more effective inter-comparison of results between studies. This was accomplished by artificially degrading a very high-quality benchmark dataset of a coastal cliff and a landslide in Norfolk, UK. Results revealed that the level of noise, image compression and downsampling all degrade the quality of 3D products from the SfM workflow. Secondly, these sets of images were pre-processed to establish whether results could be augmented by controlling the single colour channel used during photogrammetric processing. Results showed slight variations in the products generated, with evidence supporting the fine sensitivity SfM has for refining the focal length estimation of the lens. For extremely specific contexts, pre-processing of the RGB-to-single channel conversion may be relevant, but for the datasets analysed in this contribution this was not the case. Thirdly, image network configurations were investigated to build on previous research in establishing best practice. Results show that, in situations where the number of images being acquired is a limiting factor, networks with narrowly oblique overlapping images have a higher density and lower error than those with widely oblique images and those directly facing the surface normal. These results demonstrate the value of optimising image acquisition, and in the handling of this imagery. The differences in image quality and pre-processing which are unreported within geoscientific studies using SfM could account for differences between accuracies obtained, independent of the specific photogrammetric methods used. Insights from this work into how best to capture, process and produce the best quality SfM data will allow the community to adopt these best practices in the future.

The objective of this project is to develop a shape-based image analysis system, which allows classification and identification of holographic images of marine particles and plankton recorded by an underwater digital holographic camera. In order to achieve this goal, the first step is to extract shape regions of objects from images and to describe the regions by polygonal boundaries. After extraction of the polygonal boundary curve of an object, affine-invariant curve normalisation is implemented on the curve to reduce the influence of object shape deformations on object identification and classification. Six numeric features are then selected to describe shape properties of an object. Before these six shape features are used as a numeric interpretation of an object for image analysis, some processing of them is necessary, consisting of selecting the number of items in each feature and rescaling the selected feature vectors. Afterwards, Gaussian rescaling is adopted to rescale the feature data. Lastly, a shape-based image classification and identification system is built. The system contains two components: semi-automatic image classification (imCLASS) and automatic image identification (imIDENT). In imCLASS, an image retrieval method based on the support vector machine with a feedback mechanism has been developed. The function of imCLASS is to classify given images into different folders with the corresponding labels from the user. These labelled folders can be used to train the artificial neural network in imIDENT. A set of analyses of effects of the proposed methods in the process chain on image analysis are carried out. The whole performance of the system for classifying and identifying marine particles and plankton is also evaluated in terms of the time-cost and accuracy performance. In the end, some main conclusions are listed. The areas of weakness of the system are also highlighted for future work.

A vast majority of electronic systems including medical, surveillance and critical infrastructure employs image processing to provide intelligent analysis. They use onboard pre-processing to reduce data bandwidth and memory requirements before sending information to the central system. Field Programmable Gate Arrays (FPGAs) represent a strong platform as they permit reconfigurability and pipelining for streaming applications. However, rapid advances and changes in these application use cases crave adaptable hardware architectures that can process dynamic data workloads and be easily programmed to achieve ecient solutions in terms of area, time and power. FPGA-based development needs iterative design cycles, hardware synthesis and place-and-route times which are alien to the software developers. This work proposes an FPGA-based programmable hardware acceleration approach to reduce design effort and time. This allows developers to use FPGAs to profile, optimise and quickly prototype algorithms using a more familiar software-centric, edit-compile-run design flow that enables the programming of the platform by software rather than high-level synthesis (HLS) engineering principles. Central to the work has been the development of an optimised FPGA-based processor called Image Processing Processor (IPPro) which efficiently uses the underlying resources and presents a programmable environment to the programmer using a dataflow design principle. This gives superior performance when compared to competing alternatives. From this, a three-layered platform has been created which enables the realisation of parallel computing skeletons on FPGA which are used to eciently express designs in high-level programming languages. From bottom-up, these layers represent programming (actor, multiple actors and parallel skeletons) and hardware (IPPro core, multicore IPPro, system infrastructure) abstraction. The platform allows acceleration of parallel and non-parallel dataflow applications. A set of point and area image pre-processing functions are implemented on Avnet Zedboard platform which allows the evaluation of the performance. The point function achieved 2.53 times better performance than the area functions and point and area functions achieved performance improvements of 7.80 and 5.27 times over sin- gle core IPPro by exploiting data parallelism. The pipelined execution of multiple stages revealed that a dataflow graph can be decomposed into balanced actors to deliver maximum performance by hiding data transfer and processing time through exploiting task parallelism; otherwise, the maximum achievable performance is limited by the slowest actor due to the ripple effect caused by unbalanced actors. The platform delivered better performance in terms of fps/Watt/Area than Embedded Graphic Processing Unit (GPU) considering both technologies allows a software-centric design flow.

In three dimensional (3D) bone tissue engineering, the nutrient transfer process is one of the main issues that should be resolved so as to avoid the lack of nutrients (e.g., glucose and oxygen) feeding the cells in extracapillary space (ECS) region, also known as scaffold in an in vitro static culture. For example, hollow fibre membrane bioreactor (HFMB) has been built to resolve this diffusion limitation and make it possible to produce 3D bone tissues under laboratory conditions. However, in previous study, the glucose diffusivity in CCM is assumed to be same as in water, leading to inaccurate prediction of nutrient transport process in HFMB mathematical models. Hence, the overall aim of this study is to analyse and link the fluid spreading performance and glucose diffusivity across both commercial and in-house membranes and scaffolds saturated with water (reference fluid) and cell culture media (CCM) in HFMB with their microstructure with a view to carry out more accurate modelling of nutrient transport in HFMB. As a critical role in HFMB, various types of membranes and scaffolds have been used in this project, including commercial scaffolds and membranes (cellulose nitrate membrane (CN), polyvinylidene fluoride (PVDF) membrane, polycaprolactone (PCL) scaffold, collagen scaffold and poly-l-lactide acid (PLLA) scaffold), self-made 3D printed and electrospun scaffolds. The pore-morphologies of all scaffolds are analysed by scanning electronic microscope (SEM) images with a view to investigate their effects on glucose diffusivity. 3D printed scaffolds are examined in this study, as these produced scaffolds have uniform microstructural properties and precise pore morphologies. However, despite the small pore size of these scaffolds is suitable for bone tissue, their level of uniformity and regularity was found to be low. This is one of the main reasons why the scaffolds are made with electrospinning technique in this study. In the electrospinning process, the effects of operating parameters on fiber-fiber space and fiberdiameter of polycaprolactone (PCL) scaffolds on scaffolds morphologies are investigated. Given that the electrospinning process is repeatable, the fiber-fiber space and fiber diameter can be controlled, both of which increase with higher flow rate. Wettability of the tissue engineering scaffold is a critical parameter for cell attachment. But, there are limited data about the relationship between scaffolds morphology characteristics and spreading behaviour, which is focused in this study. In this study, Kruss DSA100 is used to measure the contact angle of water/ cell culture media (CCM) drop on both dry/pre-wetted varying pore size 3D printed scaffolds and electrospun PCL scaffolds. The results show that all scaffolds have low wettability. With increasing fiber-fiber space of scaffolds, the static contact angle decreases. Furthermore, compared to water drop, the contact angle is higher and it takes longer time to achieve balance with CCM drops, which show worse wettability, due to relative higher viscosity. Pre-wetted scaffolds shows better results with significantly reduced static contact angle owing to lower surface roughness and surface tension between wet surface and droplet. Hence, a conclusion can be drawn that wetting of electrospun PCL scaffolds is a prerequisite for attachment of cells. The glucose diffusivities of the scaffolds in both water (as reference) and cell culture media (CCM) with and without all membrane and scaffolds are measured, where pore size distribution, porosity and tortuosity are determined and correlated to the effective diffusivity, using a diaphragm cell method. As expected, the effective diffusivity increases correspondingly with the pore size of the materials. It is also observed that the effective glucose diffusivity through the pores of these materials in CCM is smaller than in water. It is well recognized however that it is impossible to obtain results from cell culture experiments due to its sensitivity, which makes it prominent to use other methods to predict these results. For this reason, effective glucose diffusivity is determined to predict in CCM and water by processing scanning electronic microscope (SEM) images. It can be seen that results obtained from image processing are in agreement with the experimentally obtained diffusion coefficient results as well as its porosity values. With image processed and experimtenal diffusivities, the glucose transport process in HFMB is simulated based on the Krogh cylinder assumption. The simulation shows diffusion process across membrane is the limiting step of mass transfer in HFMB. The minimum glucose concentration in the reactor is also investigated, which is found a linear relationship with glucose diffusivity across scaffold. The diffusivity across porous material measured in this study optimise the modelling and help us to have a better understanding of nutrient transfer in HFMB.

Electroluminescence (EL) imaging is a fast and comparatively low-cost method for spatially resolved analysis of photovoltaic (PV) devices. A Silicon CCD or InGaAs camera is used to capture the near infrared radiation, emitted from a forward biased PV device. EL images can be used to identify defects, like cracks and shunts but also to map physical parameters, like series resistance. The lack of suitable image processing routines often prevents automated and setup-independent quantitative analysis. This thesis provides a tool-set, rather than a specific solution to address this problem. Comprehensive and novel procedures to calibrate imaging systems, to evaluate image quality, to normalize images and to extract features are presented. For image quality measurement the signal-to-noise ratio (SNR) is obtained from a set of EL images. Its spatial average depends on the size of the background area within the EL image. In this work the SNR will be calculated spatially resolved and as (background independent) averaged parameter using only one EL image and no additional information of the imaging system. This thesis presents additional methods to measure image sharpness spatially resolved and introduces a new parameter to describe resolvable object size. This allows equalising images of different resolutions and of different sharpness allowing artefact-free comparison. The flat field image scales the emitted EL signal to the detected image intensity. It is often measured through imaging a homogeneous light source such as a red LCD screen in close distance to the camera lens. This measurement however only partially removes vignetting the main contributor to the flat field. This work quantifies the vignetting correction quality and introduces more sophisticated vignetting measurement methods. Especially outdoor EL imaging often includes perspective distortion of the measured PV device. This thesis presents methods to automatically detect and correct for this distortion. This also includes intensity correction due to different irradiance angles. Single-time-effects and hot pixels are image artefacts that can impair the EL image quality. They can conceivably be confused with cell defects. Their detection and removal is described in this thesis. The methods presented enable direct pixel-by-pixel comparison for EL images of the same device taken at different measurement and exposure times, even if imaged by different contractors. EL statistics correlating cell intensity to crack length and PV performance parameters are extracted from EL and dark I-V curves. This allows for spatially resolved performance measurement without the need for laborious flash tests to measure the light I-V- curve. This work aims to convince the EL community of certain calibration- and imaging routines, which will allow setup independent, automatable, standardised and therefore comparable results. Recognizing the benefits of EL imaging for quality control and failure detection, this work paves the way towards cheaper and more reliable PV generation. The code used in this work is made available to public as library and interactive graphical application for scientific image processing.

Confocal microscopy demonstrated its value in the diagnosis of Acanthamoeba and fungal keratitis which considered sight-threatening corneal diseases. However, it can be difficult to find and train confocal microscopy graders to accurately detect Acanthamoeba cysts and fungal filaments in the images. Use of an automated system could overcome this problem and help to start the correct treatment more quickly. Also, response to treatment can be difficult to assess in infectious keratitis using clinical examination alone, but there is evidence that the morphology of filaments and cysts may change over time with the use of correct treatment. An automated system to analyse confocal microscopy images for such changes would also assist clinicians in determining whether the ulcer is improving, or whether a change of treatment is needed. This research proposes a fully automated novel system with GUI to detect cysts and hyphae (filaments) and measure useful quantitative parameters for them through many stages; Image enhancement, image segmentation, quantitative analysis for detected cysts and hyphae, and registration and tracking of ordered sequence of images. The performance of the proposed segmentation procedure is evaluated by comparing between the manual and the automated traced images of the dataset that was provided by the Manchester Royal Eye Hospital. The positive predictive values rate of cysts for Acanthamoeba images was 76%. For detected hyphae in Fusarium images, many standard measurements were computed. The accuracy of their values was quantified by calculating the percent error rate for each measurement and which ranged from 23% to 49%.

In this research work, we develop an obstacle detection and emergency exit sign recognition system on a mobile phone by extending the feature from accelerated segment test detector with Harris corner filter. The first step often required for many vision based applications is the detection of objects of interest in an image. Hence, in this research work, we introduce emergency exit sign detection method using colour histogram. The hue and saturation component of an HSV colour model are processed into features to build a 2D colour histogram. We backproject a 2D colour histogram to detect emergency exit sign from a captured image as the first task required before performing emergency exit sign recognition. The result of classification shows that the 2D histogram is fast and can discriminate between objects and background with accuracy. One of the challenges confronting object recognition methods is the type of image feature to compute. In this work therefore, we present two feature detectors and descriptor methods based on the feature from accelerated segment test detector with Harris corner filter. The first method is called Upright FAST-Harris and binary detector (U-FaHB), while the second method Scale Interpolated FAST-Harris and Binary (SIFaHB). In both methods, feature points are extracted using the accelerated segment test detectors and Harris filter to return the strongest corner points as features. However, in the case of SIFaHB, the extraction of feature points is done across the image plane and along the scale-space. The modular design of these detectors allows for the integration of descriptors of any kind. Therefore, we combine these detectors with binary test descriptor like BRIEF to compute feature regions. These detectors and the combined descriptor are evaluated using different images observed under various geometric and photometric transformations and the performance is compared with other detectors and descriptors. The results obtained show that our proposed feature detector and descriptor method is fast and performs better compared with other methods like SIFT, SURF, ORB, BRISK, CenSurE. Based on the potential of U-FaHB detector and descriptor, we extended it for use in optical flow computation, which we termed the Nearest-flow method. This method has the potential of computing flow vectors for use in obstacle detection. Just like any other new methods, we evaluated the Nearest flow method using real and synthetic image sequences. We compare the performance of the Nearest-flow with other methods like the Lucas and Kanade, Farneback and SIFT-flow. The results obtained show that our Nearest-flow method is faster to compute and performs better on real scene images compared with the other methods. In the final part of this research, we demonstrate the application potential of our proposed methods by developing an obstacle detection and exit sign recognition system on a camera phone and the result obtained shows that the methods have the potential to solve this vision based object detection and recognition problem.

Facial expression recognition (FER) is a research area that consists of classifying the human emotions through the expressions on their face. It can be used in applications such as biometric security, intelligent human-computer interaction, robotics, and clinical medicine for autism, depression, pain and mental health problems. This dissertation investigates the advanced technologies for facial expression analysis and develops the artificial intelligent systems for practical applications. The first part of this work applies geometric and texture domain feature extractors along with various machine learning techniques to improve FER. Advanced 2D and 3D facial processing techniques such as Edge Oriented Histograms (EOH) and Facial Mesh Distances (FMD) are then fused together using a framework designed to investigate their individual and combined domain performances. Following these tests, the face is then broken down into facial parts using advanced facial alignment and localising techniques. Deep learning in the form of Convolutional Neural Networks (CNNs) is also explored also FER. A novel approach is used for the deep network architecture design, to learn the facial parts jointly, showing an improvement over using the whole face. Joint Bayesian is also adapted in the form of metric learning, to work with deep feature representations of the facial parts. This provides a further improvement over using the deep network alone. Dynamic emotion content is explored as a solution to provide richer information than still images. The motion occurring across the content is initially captured using the Motion History Histogram descriptor (MHH) and is critically evaluated. Based on this observation, several improvements are proposed through extensions such as Average Spatial Pooling Multi-scale Motion History Histogram (ASMMHH). This extension adds two modifications, first is to view the content in different spatial dimensions through spatial pooling; influenced by the structure of CNNs. The other modification is to capture motion at different speeds. Combined, they have provided better performance over MHH, and other popular techniques like Local Binary Patterns - Three Orthogonal Planes (LBP-TOP). Finally, the dynamic emotion content is observed in the feature space, with sequences of images represented as sequences of extracted features. A novel technique called Facial Dynamic History Histogram (FDHH) is developed to capture patterns of variations within the sequence of features; an approach not seen before. FDHH is applied in an end to end framework for applications in Depression analysis and evaluating the induced emotions through a large set of video clips from various movies. With the combination of deep learning techniques and FDHH, state-of-the-art results are achieved for Depression analysis.

Malignant melanoma is one of the most fatal forms of skin cancer. It has also become increasingly common, especially among white-skinned people exposed to the sun. Early detection of melanoma is essential to raise survival rates, since its detection at an early stage can be helpful and curable. Working out the dermoscopic clinical features (pigment network and lesion borders) of melanoma is a vital step for dermatologists, who require an accurate method of reaching the correct clinical diagnosis, and ensure the right area receives the correct treatment. These structures are considered one of the main keys that refer to melanoma or non-melanoma disease. However, determining these clinical features can be a time-consuming, subjective (even for trained clinicians) and challenging task for several reasons: lesions vary considerably in size and colour, low contrast between an affected area and the surrounding healthy skin, especially in early stages, and the presence of several elements such as hair, reflections, oils and air bubbles on almost all images. This thesis aims to provide an accurate, robust and reliable automated dermoscopy image analysis technique, to facilitate the early detection of malignant melanoma disease. In particular, four innovative methods are proposed for region segmentation and classification, including two for pigmented region segmentation, one for pigment network detection, and one for lesion classification. In terms of boundary delineation, four pre-processing operations, including Gabor filter, image sharpening, Sobel filter and image inpainting methods are integrated in the segmentation approach to delete unwanted objects (noise), and enhance the appearance of the lesion boundaries in the image. The lesion border segmentation is performed using two alternative approaches. The Fuzzy C-means and the Markov Random Field approaches detect the lesion boundary by repeating the labeling of pixels in all clusters, as a first method. Whereas, the Particle Swarm Optimization with the Markov Random Field method achieves greater accuracy for the same aim by combining them in the second method to perform a local search and reassign all image pixels to its cluster properly. With respect to the pigment network detection, the aforementioned pre-processing method is applied, in order to remove most of the hair while keeping the image information and increase the visibility of the pigment network structures. Therefore, a Gabor filter with connected component analysis are used to detect the pigment network lines, before several features are extracted and fed to the Artificial Neural Network as a classifier algorithm. In the lesion classification approach, the K-means is applied to the segmented lesion to separate it into homogeneous clusters, where important features are extracted; then, an Artificial Neural Network with Radial Basis Functions is trained by representative features to classify the given lesion as melanoma or not. The strong experimental results of the lesion border segmentation methods including Fuzzy C-means with Markov Random Field and the combination between the Particle Swarm Optimization and Markov Random Field, achieved an average accuracy of 94.00% , 94.74% respectively. Whereas, the lesion classification stage by using extracted features form pigment network structures and segmented lesions achieved an average accuracy of 90.1% , 95.97% respectively. The results for the entire experiment were obtained using a public database PH2 comprising 200 images. The results were then compared with existing methods in the literature, which have demonstrated that our proposed approach is accurate, robust, and efficient in the segmentation of the lesion boundary, in addition to its classification.

Unmanned aerial vehicles (UAV) have become an interesting and active research topic for photogrammetry. Current research is based on images acquired by an UAV, which have a high ground resolution and good spectral and radiometrical resolution, due to the low flight altitudes combined with a high resolution camera. UAV image flights are also cost effective and have become attractive for many applications including, change detection in small scale areas. One of the main problems preventing full automation of data processing of UAV imagery is the degradation effect of blur caused by camera movement during image acquisition. This can be caused by the normal flight movement of the UAV as well as strong winds, turbulence or sudden operator inputs. This blur disturbs the visual analysis and interpretation of the data, causes errors and can degrade the accuracy in automatic photogrammetric processing algorithms. The detection and removal of these images is currently achieved manually, which is both time consuming and prone to error, particularly for large image-sets. To increase the quality of data processing an automated process is necessary, which must be both reliable and quick. This thesis proves the negative affect that blurred images have on photogrammetric processing. It shows that small amounts of blur do have serious impacts on target detection and that it slows down processing speed due to the requirement of human intervention. Larger blur can make an image completely unusable and needs to be excluded from processing. To exclude images out of large image datasets an algorithm was developed. The newly developed method makes it possible to detect blur caused by linear camera displacement. The method is based on human detection of blur. Humans detect blurred images best by comparing it to other images in order to establish whether an image is blurred or not. The developed algorithm simulates this procedure by creating an image for comparison using image processing. Creating internally a comparable image makes the method independent of additional images. However, the calculated blur value named SIEDS (saturation image edge difference standard-deviation) on its own does not provide an absolute number to judge if an image is blurred or not. To achieve a reliable judgement of image sharpness the SIEDS value has to be compared to other SIEDS values of the same dataset. This algorithm enables the exclusion of blurred images and subsequently allows photogrammetric processing without them. However, it is also possible to use deblurring techniques to restor blurred images. Deblurring of images is a widely researched topic and often based on the Wiener or Richardson-Lucy deconvolution, which require precise knowledge of both the blur path and extent. Even with knowledge about the blur kernel, the correction causes errors such as ringing, and the deblurred image appears muddy and not completely sharp. In the study reported in this paper, overlapping images are used to support the deblurring process. An algorithm based on the Fourier transformation is presented. This works well in flat areas, but the need for geometrically correct sharp images for deblurring may limit the application. Another method to enhance the image is the unsharp mask method, which improves images significantly and makes photogrammetric processing more successful. However, deblurring of images needs to focus on geometric correct deblurring to assure geometric correct measurements. Furthermore, a novel edge shifting approach was developed which aims to do geometrically correct deblurring. The idea of edge shifting appears to be promising but requires more advanced programming.

Airborne multi-sensor imaging is increasingly used to examine vegetation properties. The advantage of using multiple types of sensor is that each detects a different feature of the vegetation, so that collectively they provide a detailed understanding of the ecological pattern. Specifically, Light Detection And Ranging (LiDAR) devices produce detailed point clouds of where laser pulses have been backscattered from surfaces, giving information on vegetation structure; hyperspectral sensors measure reflectances within narrow wavebands, providing spectrally detailed information about the optical properties of targets; while aerial photographs provide high spatial-resolution imagery so that they can provide more feature details which cannot be identified from hyperspectral or LiDAR intensity images. Using a combination of these sensors, effective techniques can be developed for mapping species and inferring leaf physiological processes at ITC-level. Although multi-sensor approaches have revolutionised ecological research, their application in mapping individual tree crowns is limited by two major technical issues: (a) Multi-sensor imaging requires all images taken from different sensors to be co-aligned, but different sensor characteristics result in scale, rotation or translation mismatches between the images, making correction a pre-requisite of individual tree crown mapping; (b) reconstructing individual tree crowns from unstructured raw data space requires an accurate tree delineation algorithm. This thesis develops a schematic way to resolve these technical issues using the-state-of-the-art computer vision algorithms. A variational method, called NGF-Curv, was developed to co-align hyperspectral imagery, LiDAR and aerial photographs. NGF-Curv algorithm can deal with very complex topographic and lens distortions efficiently, thus improving the accuracy of co-alignment compared to established image registration methods for airborne data. A graph cut method, named MCNCP-RNC was developed to reconstruct individual tree crowns from fully integrated multi-sensor imagery. MCNCP-RNC is not influenced by interpolation artefacts because it detects trees in 3D, and it detects individual tree crowns using both hyperspectral imagery and LiDAR. Based on these algorithms, we developed a new workflow to detect species at pixel and ITC levels in a temperate deciduous forest in the UK. In addition, we modified the workflow to monitor physiological responses of two oak species with respect to environmental gradients in a Mediterranean woodland in Spain. The results show that our scheme can detect individual tree crowns, find species and monitor physiological responses of canopy leaves.