Search

The phenomenon of Adversarial Examples has become one of the most intriguing topics associated to deep learning. The so-called adversarial attacks have the ability to fool deep neural networks with inappreciable perturbations. While the effect is striking, it has been suggested that such carefully selected injected noise does not necessarily appear in real-world scenarios. In contrast to this, some authors have looked for ways to generate adversarial noise in physical scenarios (traffic signs, shirts, etc.), thus showing that attackers can indeed fool the networks. In this paper we go beyond that and show that adversarial examples also appear in the real-world without any attacker or maliciously selected noise involved. We show this by using images from tasks related to microscopy and also general object recognition with the well-known ImageNet dataset. A comparison between these natural and the artificially generated adversarial examples is performed using distance metrics and image quality metrics. We also show that the natural adversarial examples are in fact at a higher distance from the originals that in the case of artificially generated adversarial examples.

Fast automatic handgun detection can be very useful to avoid or mitigate risks in public spaces. Detectors based on deep learning methods have been proposed in the literature to trigger an alarm if a handgun is detected in the image. However, those detectors are solely based on the weapon appearance on the image. In this work, we propose to combine the detector with the individual’s pose information in order to improve overall performance. To this end, a model that integrates grayscale images from the output of the handgun detector and heatmap-like images that represent pose is proposed. The results show an improvement over the original handgun detector. The proposed network provides a maximum improvement of a 17.5% in AP of the proposed combinational model over the baseline handgun detector.

In an object detection system, the main objective during training is to maintain the detection and false positive rates under acceptable levels when the model is run over the test set. However, this typically translates into an unacceptable rate of false alarms when the system is deployed in a real surveillance scenario. To deal with this situation, which often leads to system shutdown, we propose to add a filter step to discard part of the new false positive detections that are typical of the new scenario. This step consists of a deep autoencoder trained with the false alarm detections generated after running the detector over a period of time in the new scenario. Therefore, this step will be in charge of determining whether the detection is a typical false alarm of that scenario or whether it is something anomalous for the autoencoder and, therefore, a true detection. In order to decide whether a detection must be filtered, three different approaches have been tested. The first one uses the autoencoder reconstruction error measured with the mean squared error to make the decision. The other two use the k-NN (k-nearest neighbors) and one-class SVMs (support vector machines) classifiers trained with the autoencoder vector representation. In addition, a synthetic scenario has been generated with Unreal Engine 4 to test the proposed methods in addition to a dataset with real images. The results obtained show a reduction in the number of false positives between 22.5% and 87.2% and an increase in the system’s precision of 1.2%$$-47$$ - 47 % when the autoencoder is applied.

Deep learning (henceforth DL) has become most powerful machine learning methodology. Under specific circumstances recognition rates even surpass those obtained by humans. Despite this, several works have shown that deep learning produces outputs that are very far from human responses when confronted with the same task. This the case of the so-called “adversarial examples” (henceforth AE). The fact that such implausible misclassifications exist points to a fundamental difference between machine and human learning. This paper focuses on the possible causes of this intriguing phenomenon. We first argue that the error in adversarial examples is caused by high bias, i.e. by regularization that has local negative effects. This idea is supported by our experiments in which the robustness to adversarial examples is measured with respect to the level of fitting to training samples. Higher fitting was associated to higher robustness to adversarial examples. This ties the phenomenon to the trade-off that exists in machine learning between fitting and generalization.

[Abstract] Unfortunately, attacks with firearms such as handguns have become too common. CCTV surveillance systems can potentially help to prevent this kind of incidents, but require continuous human supervision, which is not feasible in practice. Image-based handgun detectors allow the automatic location of these weapons to send alerts to the security staff. Deep learning has been recently used for this purpose. However, the precision and sensitivity of these systems are not generally satisfactory, causing in most cases both false alarms and undetected handguns, particularly when the firearm is far from the camera. This paper proposes the use of information related to the pose of the subject to improve the performance of current handgun detectors. More concretely, a human full-body pose classifier has been developed which is capable of separating between shooting poses and other non-dangerous poses. The classified pose is then used to reduce both the number of false positives (FP) and false negatives (FN). The proposed method has been tested with several datasets and handgun detectors, showing an improvement under various metrics. This work was partially funded by projects PDC2021-121197-C22 (funded by MCIN/AEI/ 10.13039/501100011033 and by the European Union NextGenerationEU/PRTR) and SBPLY/21/180501/000025 (funded by the Autonomous Government of Castilla-La Mancha and the European Regional Development Fund -ERDF-). The first author is supported by Postgraduate Grant PRE2018-083772 from the Spanish Ministry of Science, Innovation, and Universities. Junta de Comunidades de Castilla-La Mancha; SBPLY/21/180501/000025

Training a deep convolutional neural network (CNN) for automatic image classification requires a large database with images of labeled samples. However, in some applications such as biology and medicine only a few experts can correctly categorize each sample. Experts are able to identify small changes in shape and texture which go unnoticed by untrained people, as well as distinguish between objects in the same class that present drastically different shapes and textures. This means that currently available databases are too small and not suitable to train deep learning models from scratch. To deal with this problem, data augmentation techniques are commonly used to increase the dataset size. However, typical data augmentation methods introduce artifacts or apply distortions to the original image, which instead of creating new realistic samples, obtain basic spatial variations of the original ones.We propose a novel data augmentation procedure which generates new realistic samples, by combining two samples that belong to the same class. Although the idea behind the method described in this paper is to mimic the variations that diatoms experience in different stages of their life cycle, it has also been demonstrated in glomeruli and pollen identification problems. This new data augmentation procedure is based on morphing and image registration methods that perform diffeomorphic transformations.The proposed technique achieves an increase in accuracy over existing techniques of 0.47%, 1.47%, and 0.23% for diatom, glomeruli and pollen problems respectively.For the Diatom dataset, the method is able to simulate the shape changes in different diatom life cycle stages, and thus, images generated resemble newly acquired samples with intermediate shapes. In fact, the other methods compared obtained worse results than those which were not using data augmentation. For the Glomeruli dataset, the method is able to add new samples with different shapes and degrees of sclerosis (through different textures). This is the case where our proposed DA method is more beneficial, when objects highly differ in both shape and texture. Finally, for the Pollen dataset, since there are only small variations between samples in a few classes and this dataset has other features such as noise which are likely to benefit other existing DA techniques, the method still shows an improvement of the results.

Introducing efficient automatic violence detection in video surveillance or audiovisual content monitoring systems would greatly facilitate the work of closed-circuit television (CCTV) operators, rating agencies or those in charge of monitoring social network content. In this paper we present a new deep learning architecture, using an adapted version of DenseNet for three dimensions, a multi-head self-attention layer and a bidirectional convolutional long short-term memory (LSTM) module, that allows encoding relevant spatio-temporal features, to determine whether a video is violent or not. Furthermore, an ablation study of the input frames, comparing dense optical flow and adjacent frames subtraction and the influence of the attention layer is carried out, showing that the combination of optical flow and the attention mechanism improves results up to 4.4%. The conducted experiments using four of the most widely used datasets for this problem, matching or exceeding in some cases the results of the state of the art, reducing the number of network parameters needed (4.5 millions), and increasing its efficiency in test accuracy (from 95.6% on the most complex dataset to 100% on the simplest one) and inference time (less than 0.3 s for the longest clips). Finally, to check if the generated model is able to generalize violence, a cross-dataset analysis is performed, which shows the complexity of this approach: using three datasets to train and testing on the remaining one the accuracy drops in the worst case to 70.08% and in the best case to 81.51%, which points to future work oriented towards anomaly detection in new datasets. Ministerio de Economía y Competitividad TIN2017-82113-C2-1-R MInisterio de Economía y Competitividad TIN2017-82113-C2-2-R

[Resumen] Los microscopios son una herramienta imprescindible para diferentes aplicaciones en biología y medicina, entre otros. La automatización de estos microscopios y la integración de herramientas de análisis automático de imagen es un gran avance en microscopía, pero la mayoría de los sistemas disponibles en el mercado con estas prestaciones tienen un precio elevado. Por otro lado, en los últimos años ha habido un auge importante del desarrollo y venta de impresoras 3D a un precio muy asequible. Por ello, en este trabajo se busca ofrecer un kit modular que permita convertir una impresora 3D en un microscopio motorizado completamente funcional y automatizado. Para conseguir este objetivo, se ha seleccionado una impresora 3D, de la que se ha reaprovechado el eje de movimiento cartesiano, y se le ha acoplado un sistema óptico basado en Raspberry Pi HQ Camera, una iluminación superior e inferior y además de una base para portaobjetos nueva. Así como herramientas de procesado y métodos de clasificación automática de la imagen. Con esto se conseguirá un sistema de microscopía modular y de bajo coste, capaz de capturar muestras microscópicas de organismos y tejidos en portas 2D y en placas Petri 3D. [Abstract] Microscopes are an essential tool for different applications in biology and medicine, among others. The automation of these microscopes and the integration of automatic image analysis tools is a breakthrough in microscopy. However, most of the systems available on the market with these features have a high price, which prevents them from becoming more popular. On the other hand, in recent years there has been a significant boom in the development and sale of 3D printers at a very affordable price. Therefore, this work seeks to offer a modular kit that allows converting a 3D printer into a fully functional and automated microscope. To achieve this goal, all possible parts of a commercially available 3D printer will be used, to which all the necessary optical and lighting elements will be attached. This will result in a modular and cost-effective microscopy system. Junta de Comunidades de Castilla-La Mancha;SBPLY/19/180501/00 0273

The data presented in this article is part of the whole slide imaging (WSI) datasets generated in European project AIDPATH 2 This data is also related to the research paper entitle “Glomerulosclerosis Identification in Whole Slide Images using Semantic Segmentation”, published in Computer Methods and Programs in Biomedicine Journal [1]. In that article, different methods based on deep learning for glomeruli segmentation and their classification into normal and sclerotic glomerulous are presented and discussed. The raw data used is described and provided here. In addition, the detected glomeruli are also provided as individual image files. These data will encourage research on artificial intelligence (AI) methods, create and compare fresh algorithms, and measure their usability in quantitative nephropathology. Keywords: Glomeruli identification, Normal glomerulus, Global sclerotic glomerulus, Whole slide image, Digital pathology

21 pags., 11 figs., 7 tabs., A microscope is an essential tool in biosciences and production quality laboratories for unveiling the secrets of microworlds. This paper describes the development of MicroHikari3D, an affordable DIY optical microscopy platform with automated sample positioning, autofocus and several illumination modalities to provide a high-quality flexible microscopy tool for labs with a short budget. This proposed optical microscope design aims to achieve high customization capabilities to allow whole 2D slide imaging and observation of 3D live specimens. The MicroHikari3D motion control system is based on the entry level 3D printer kit Tronxy X1 controlled from a server running in a Raspberry Pi 4. The server provides services to a client mobile app for video/image acquisition, processing, and a high level classification task by applying deep learning models., Junta de Comunidades de Castilla-La Mancha (SBPLY/19/180501/000273).

While action recognition has become an important line of research in computer vision, the recognition of particular events such as aggressive behaviors, or fights, has been relatively less studied. These tasks may be exceedingly useful in some video surveillance scenarios such as psychiatric centers, prisons or even in personal camera smartphones. Their potential usability has caused a surge of interest in developing fight or violence detectors. The key aspect in this case is efficiency, that is, these methods should be computationally very fast. In this paper, spatio-temporal elastic cuboid trajectories are proposed for fight recognition. This method is based on the use of blob movements to create trajectories that capture and model the different motions that are specific to a fight. The proposed method is robust to the specific shapes and positions of the individuals. Additionally, the standard Hough forests classifier is adapted in order to use it with this descriptor. This method is compared to other nine related methods on four datasets. The results show that the proposed method obtains the best accuracy for each dataset and is also computationally efficient.

This chapter presents the most relevant image processing techniques and algorithms related to computing features that are able to characterize diatoms as objects in the computer vision field for further analysis and classification. For this purpose, a wide revision of the most important contributions to diatom classification is performed. Moreover, features that have been found to be suitable for this task are covered. Later on, the reader will find the main techniques for diatom classification for the two paradigms that are used nowadays: machine learning with classical methods that rely on previously selected features, or deep learning, which learns the features from the images automatically.

15 pags., 21 figs., 4 tabs., Microscopic algae segmentation, specifically of diatoms, is an essential procedure for water quality assessment. The segmentation of these microalgae is still a challenge for computer vision. This paper addresses for the first time this problem using deep learning approaches to predict exactly those pixels that belong to each class, i.e., diatom and non diatom. A comparison between semantic segmentation and instance segmentation is carried out, and the performance of these methods is evaluated in the presence of different types of noise. The trained models are then evaluated with the same raw images used for manual diatom identification. A total of 126 images of the entire field of view at 60x magnification, with a size of 2592x1944 pixels, are analyzed. The images contain 10 different taxa plus debris and fragments. The best results were obtained with instance segmentation achieving an average precision of 85% with 86% sensitivity and 91% specificity (up to 92% precision with 98%, both sensitivity and specificity for some taxa). Semantic segmentation was able to improve the average sensitivity up to 95% but decreasing the specificity down to 60% and precision to 57%. Instance segmentation was also able to properly separate diatoms when overlap occurs, which helps estimate the number of diatoms, a key requirement for water quality grading., The authors received financial support from the Spanish Government under the Aqualitas-retos project (Ref. CTM2014-51907-C2-2-R-MINECO).

IbPRIA 2019,Madrid, Spain. July 1-4 2019. -- http://www.ibpria.org/2019/. -- Part of the Lecture Notes in Computer Science book series (LNCS, volume 11867) Also part of the Image Processing, Computer Vision, Pattern Recognition, and Graphics book sub series (LNIP, volume 11867), Diatom identification is a crucial process to estimate water quality, which is essential in biological studies. This process can be auto- mated with machine learning algorithms. For this purpose, a dataset with 10 common taxa is collected, with annotations provided by an expert diatomist. In this work, a comparison of the classical state-of-the-art gen- eral purpose methods along with two different deep learning approaches is carried out. The classical methods are based on Viola-Jones and scale and curvature invariant ridge object detectors. The deep learning based methods are Semantic Segmentation and YOLO. This is the first time that Viola-Jones and Semantic Segmentation techniques are applied and compared for diatom segmentation in microscopic images containing sev- eral taxon shells. While all methods provide relatively good results in specific species, the deep learning approaches are consistently better in terms of sensitivity and specificity (up to 0.99 for some taxa) and up to 0.86 precision., The authors acknowledge financial support of the Spanish Government under the Aqualitas-retos project (Ref. CTM2014-51907-C2-R-MINECO).

In recent scientific literature, some studies have been published where recognition rates obtained with Deep Learning (DL) surpass those obtained by humans on the same task. In contrast to this, other studies have shown that DL networks have a somewhat strange behavior which is very different from human responses when confronted with the same task. The case of the so-called “adversarial examples” is perhaps the best example in this regard. Despite the biological plausibility of neural networks, the fact that they can produce such implausible misclassifications still points to a fundamental difference between human and machine learning. This paper delves into the possible causes of this intriguing phenomenon. We first contend that, if adversarial examples are pointing to an implausibility it is because our perception of them relies on our capability to recognise the classes of the images. For this reason we focus on what we call cognitively adversarial examples, which are those obtained from samples that the classifier can in fact recognise correctly. Additionally, in this paper we argue that the phenomenon of adversarial examples is rooted in the inescapable trade-off that exists in machine learning (including DL) between fitting and generalization. This hypothesis is supported by experiments carried out in which the robustness to adversarial examples is measured with respect to the degree of fitting to the training samples.

[Resumen] La demanda de datos para el entrenamiento de las nuevas t ecnicas de aprendizaje profundo se ha incrementado durante los ultimos a~nos. Aunque se ha creado una comunidad extensa alrededor del intercambio de datos, e incluso muchos de los conjuntos de datos de grandes empresas se han publicado de forma gratuita, contin ua habiendo problemas espec cos para los que no se dispone de conjuntos espec cos para el entrenamiento de los modelos que los resuelvan. Este es el caso de la detecci on de armas en escenas videovigiladas donde la detecci on temprana de situaciones y objetos peligrosos es de vital importancia. Varias han sido las soluciones propuestas en los ultimos a~nos al respecto pero la adquisici on de los datos necesarios para su desarrollo sigue siendo un problema. Por ello, en este trabajo se propone generar im agenes de videovigilancia con un motor gr a co y comprobar si estos datos sint eticos pueden sustituir la captura y el etiquetado de im agenes reales [Abstract] With the development of the new deep lear- ning techniques, the data demand for trai- ning these models has increased. Although a large community has been created around data and even big companies have relea- se their own datasets free of charge, there are speci c problems for which training da- tasets are not available. This is the case of weapon detection in video-surveillance where the early detection of dangerous si- tuations and objects is of vital importance. Several solutions have been proposed in the last years but the data barrier is still a pro- blem. Therefore, in this work we propose to generate video surveillance images with a graphical engine and check if the synt- hetic data generated can replace collecting and labeling real images Ministerio de Economía y Empresa; TIN2017-82113-C2-2-R Junta de Castilla la Mancha; SBPLY/17/180501/000543

Public safety in public areas is nowadays one of the main concerns for governments and companies around the world. Video surveillance systems can take advantage from the emerging techniques of deep learning to improve their performance and accuracy detecting possible threats. This paper presents a system for gun and knife detection based on the Faster R-CNN methodology. Two approaches have been compared taking as CNN base a GoogleNet and a SqueezeNet architecture respectively. The best result for gun detection was obtained using a SqueezeNet architecture achieving a 85.44% \(AP_{50}\). For knife detection, the GoogleNet approach achieved a 46.68% \(AP_{50}\). Both results improve upon previous literature results evidencing the effectiveness of our detectors.

Law enforcement agencies and private security companies work to prevent, detect and counteract any threat with the resources they have, including alarms and video surveillance. Even so, there are still terrorist attacks or shootings in schools in which armed people move around a venue exercising violence and generating victims, showing the limitations of current systems. For example, they force security agents to monitor continuously all the images coming from the installed cameras, and potential victims nearby are not aware of the danger until someone triggers a general alarm, which also does not give them information on what to do to protect themselves. In this article we present a project that is being developed to apply the latest technologies in early threat detection and optimal response. The system is based on the automatic processing of video surveillance images to detect weapons and a mobile app that serves both for detection through the analysis of mobile device sensors, and to send users personalised and dynamic indications. The objective is to react in the shortest possible time and minimise the damage suffered.

The development of object detection systems is normally driven to achieve both high detection and low false positive rates in a certain public dataset. However, when put into a real scenario the result is generally an unacceptable rate of false alarms. In this context we propose to add an additional step that models and filters the typical false alarms of the new scenario while roughly maintaining the ability to detect the objects of interest. We propose to use the false alarms of the new scenario to train a deep autoencoder and to model them. The latter will act as a filter that checks whether the output of the detector is one of its typical false positives or not based on the reconstruction error measured with the Mean Squared Error (MSE) and the Peak Signal-to-Noise Ratio (PSNR). We test the system using an entirely synthetic novel dataset for training and testing the autoencoder generated with Unreal Engine 4. Results show a reduction in the number of FPs of up to 37.9% in combination with the PSNR error while maintaining the same detection capability.

[Abstract] Breast biopsies are crucial in the process of detec ing a wide range of diseases such as breast cancer. The evaluation of these biopsies is performed by trained pathologists that are often overworked due to the increasing number of pathologies requested. Automatic tumour detection techniques have been developed, achieving very good results. In this work, we propose to classify breast biopsies in all the different types of tissue present in them. The tissue types were identified by hand-labeling them following the indications of an expert pathologist. Afterward, they were trained with diffeerent convolutional neural networks such as GoogleNet, AlexNet, SqueezeNet and DenseNet. Out of these four networks, GoogleNet outperformed all of them achieving 95.4% of accuracy. Finally, we tried to identify why the networks were underperforming while also suggesting how results could be improved. AIDPATH EC; 612471

Background and Objective: Glomeruli identification, i.e., detection and characterization, is a key procedure in many nephropathology studies. In this paper, semantic segmentation based on convolutional neural networks (CNN) is proposed to detect glomeruli using Whole Slide Imaging (WSI) follows by a classification CNN to divide the glomeruli into normal and sclerosed. Methods: Comparison between U-Net and SegNet CNNs is performed for pixel-level segmentation considering both a two and three class problem, that is, a) non-glomerular and glomerular structures and b) non-glomerular normal glomerular and sclerotic structures. The two class semantic segmentation result is then used for a CNN classification where glomerular regions are divided into normal and global sclerosed glomeruli. Results: These methods were tested on a dataset composed of 47 WSIs belonging to human kidney sections stained with Periodic Acid Schiff (PAS). The best approach was the SegNet for two class segmentation follows by a fine-tuned AlexNet network to characterize the glomeruli. 98.16% of accuracy was obtained with this process of consecutive CNNs (SegNet-AlexNet) for segmentation and classification. Conclusion: The results obtained demonstrate that the sequential CNN segmentation-classification strategy achieves higher accuracy reducing misclassified cases and therefore being the methodology proposed for glomerulosclerosis detection.

While action recognition has become an important line of research in computer vision, the recognition of particular events, such as aggressive behaviors, or fights, has been relatively less studied. These tasks may be extremely useful in several video surveillance scenarios, such as psychiatric wards, prisons, or even in personal camera smartphones. Their potential usability has led to a surge of interest in developing fight or violence detectors. One of the key aspects in this case is efficiency, that is, these methods should be computationally fast. “Handcrafted” spatio-temporal features that account for both motion and appearance information can achieve high accuracy rates, albeit the computational cost of extracting some of those features is still prohibitive for practical applications. The deep learning paradigm has been recently applied for the first time to this task too, in the form of a 3D convolutional neural network that processes the whole video sequence as input. However, results in human perception of other’s actions suggest that, in this specific task, motion features are crucial. This means that using the whole video as input may add both redundancy and noise in the learning process. In this paper, we propose a hybrid “handcrafted/learned” feature framework which provides better accuracy than the previous feature learning method, with similar computational efficiency. The proposed method is compared to three related benchmark data sets. The method outperforms the different state-of-the-art methods in two of the three considered benchmark data sets.