Search

We investigated optimal implementation strategies for industrial inspection systems aiming to detect cracks on ground steel billets’ surfaces by combining state-of-the-art AI-based methods and classical computational imaging techniques. In 2D texture images, the interesting patterns of surface irregularities are often surrounded by visual clutter, which is to be ignored, e.g., grinding patterns. Even neural networks struggle to reliably distinguish between actual surface disruptions and irrelevant background patterns. Consequently, the image acquisition procedure already has to be optimised to the specific application. In our case, we use photometric stereo (PS) imaging to generate 3D surface models of steel billets using multiple illumination units. However, we demonstrate that the neural networks, especially in high-speed scenarios, still suffer from recognition deficiencies when using raw photometric stereo camera data, and are unable to generalise to new billets and image acquisition conditions. Only the additional application of adequate state-of-the-art image processing algorithms guarantees the best results in both aspects. The neural networks benefit when appropriate image acquisition methods together with image processing algorithms emphasise relevant surface structures and reduce overall pattern variation. Our proposed combined strategy shows a 9.25% better detection rate on validation data and is 14.7% better on test data, displaying the best generalisation.

We present a system to extract surface orientation and albedos from a single shot image using three differently colored illumination sources. Photometric stereo allows one to extract local surface information such as normals or gradients. Traditionally, the local orientations and albedos are computed using serveral acquisitions of the same viewing angle and under varying illumination directions. In applications with moving objects, where the acquisition-as well as processing speed are essential, such setups are poorly suited. We propose a single shot decomposition using three differently colored light sources under defined illumination directions. To allow for a fast and regularized inference, we built a split U-shaped convolutional neural network, which takes a single shot input and estimates both the surface orientation and albedo simultaneously.

Generative Adversarial Networks (GANs) have been used to model the underlying probability distribution of sample based datasets. GANs are notoriuos for training difficulties and their dependence on arbitrary hyperparameters. One recent improvement in GAN literature is to use the Wasserstein distance as loss function leading to Wasserstein Generative Adversarial Networks (WGANs). Using this as a basis, we show various ways in which the Wasserstein distance is estimated for the task of generative modelling. Additionally, the secrets in training such models are shown and summarized at the end of this work. Where applicable, we extend current works to different algorithms, different cost functions, and different regularization schemes to improve generative models.

We report the case of a 19-years-old patient who presented with a perplexing variety of symptoms which included remarkable facial features, intellectual disability, granulomatous upper lip swelling (previously diagnosed as Melkersson–Rosenthal syndrome), Crohn’s-like disease, non-productive cough, and a granulomatous mass localized in the left lung. Chronic granulomatous disease (CGD) was diagnosed using dihydrorhodamine 123 assay that showed low levels of phagocytic NADPH-oxidase. DNA sequencing revealed a heterozygous mutation in the NCF-1 gene on chromosome 7. As remarkable facial features and psychomotor retardation are not associated with CGD, a more detailed genetic work-up using fluorescence in situ hybridization was performed. A microdeletion in 7q11.23 on one allele indicated Williams–Beuren syndrome (WBS). The NCF-1 gene and its two pseudogenes are part of a highly repetitive region within 7q11.23 and are prone to recombination events and deletions. Such deletions can involve both the WBS critical region and the NCF-1 wildtype gene, as was the case for our patient. The second allele of the NCF-1 gene was affected by the frequent c.75.76delGT mutation that stems from a recombination of the NCF-1 wildtype gene with one of its pseudogenes. In conclusion, patients with NCF-1-deficient CGD may also harbor microdeletions that result in WBS or other hereditary disorders; therefore, it is important to perform a thorough genetic analysis in order to initiate appropriate therapy for these patients. CA Soukup

In continental Europe, the pedagogical, institutional and networking developments in the field of creative writing lag behind the United States, Britain or Australia. The starting point for more in...

Most color image sensors use color filter arrays (CFA). With this sensor design the captured information at each sensor pixel position is restricted to a specific spectral portion (typically red, green and blue bands). To obtain the missing color responses at each pixel position, so-called CFA demosaicing algorithms are commonly used. We propose two new CFA demosaicing algorithms, which are well suited for industrial print inspection with respect to the requirements in accuracy and speed. As a main contribution, we introduce novel demosaicing algorithms for specific high-speed color digital time delay and integration (DTDI) CFA line-scan cameras. We compare the suggested CFA demosaicing algorithms to state-of-the art algorithms for area and line-scan camera operation modes. We show that the two new algorithms perform superior to conventional algorithms as indicated by reconstruction error.

This article presents an approach for finding displacements of print layers from sequences of sample images semi-automatically with the aim to simplify and shorten the setup of inspection systems for printing processes in which the perfect alignment of print layers cannot be guaranteed. The basic idea behind the proposed approach is to identify pixels which are likely to have the same displacements for a given pair of images. This relatively coarse information is computed for several pairs of sample images and aggregated in order to identify regions that tend to have the same displacement over a large proportion of image pair comparisons. This idea is motivated and justified in detail. The test cases considered in this study are data from banknote print inspection. We use these data to illustrate the steps of the algorithm. The examples demonstrate the method's capability to sensibly identify print layers, even if they overlap partially. Although the paper concentrates on a particular case study, the method can be used in any print inspection process with similar characteristics.

Diffractive Optically Variable Image Devices (DOVIDs) are popular security features used to protect security documents such as banknotes, ID cards, passports, etc. Checking authenticity of these security features on both user as well as forensic level remains a challenging task, requiring sophisticated hardware tools and expert knowledge. Recently, we proposed a technique exploiting a large-scale photometric behavior of DOVIDs in order to discriminate denominations and detect counterfeits. Here we investigate invariance properties of the proposed method and demonstrate its robustness against various common perturbations, which may have negative impact on the acquisition quality in practice. Presented results show a great potential of this approach primarily for security and forensic purposes, but also for other applications, where automated inspection of DOVIDs is of interest.

We present a light-field multi-line-scan image acquisition and processing system intended for the 2.5/3-D inspection of fine surface structures, such as small parts, security print, etc. in an industrial environment. The system consists of an area-scan camera, that allows for a small number of sensor lines to be extracted at high frame rates, and a mechanism for transporting the inspected object at a constant speed. During the acquisition, the object is moved orthogonally to the camera’s optical axis as well as the orientation of the sensor lines. In each time step, a predefined subset of lines is read out from the sensor and stored. Afterward, by collecting all corresponding lines acquired over time, a 3-D light field is generated, which consists of multiple views of the object observed from different viewing angles while transported w.r.t. the acquisition device. This structure allows for the construction of so-called epipolar plane images (EPIs) and subsequent EPI-based analysis in order to achieve two main goals: (i) the reliable estimation of a dense depth model and (ii) the construction of an all-in-focus intensity image. Beside specifics of our hardware setup, we also provide a detailed description of algorithmic solutions for the mentioned tasks. Two alternative methods for EPI-based analysis are compared based on artificial and real-world data.

Convolutional neural networks (CNNs) achieved impressive recognition rates in image classification tasks recently. In order to exploit those capabilities, we trained CNNs on a database of photometric stereo images of metal surface defects, i.e. rail defects. Those defects are cavities in the rail surface and are indication for further surface degradation right up to rail break. Due to security issues, defects have to be recognized early in order to take countermeasures in time. By means of differently colored light-sources illuminating the rail surfaces from different and constant directions, those cavities are made visible in a photometric dark-field setup. So far, a model-based approach has been used for image classification, which expressed the expected reflection properties of surface defects in contrast to non-defects. In this work, we experimented with classical CNNs trained in pure supervised manner and also explored the impact of regularization methods such as unsupervised layer-wise pre-training and training data-set augmentation. The classical CNN already distinctly outperforms the model-based approach. Moreover, regularization methods yet yield further improvements.

We present a method exploiting computational refocusing capabilities of a light-field camera in order to obtain 3D shape information. We consider a light-field constructed from the relative motion between a camera and observed objects, i.e. points on the object surface are imaged under different angles along the direction of the motion trajectory. Computationally refocused images are handled by a shape-from-focus algorithm. A linear sharpness measure is shown to be computational advantageous as computational refocusing to a specific depth and sharpness assessment of each refocused image can be reordered. We also present a view matching method which further stabilizes the suggested procedure when fused with sharpness assessment. Results for real-world objects from an inspection task are presented. Comparison to ground-truth data showed average depth errors on the order of magnitude of 1 mm for a depth range of 1 cm.

We present algorithms for depth estimation from light-field data acquired by a multi-line-scan image acquisition system. During image acquisition a 3-D light field is generated over time, which consists of multiple views of the object observed from different viewing angles. This allows for the construction of so-called epipolar plane images (EPIs) and subsequent EPI-based depth estimation. We compare several approaches based on testing various slope hypotheses in the EPI domain, which can directly be related to depth. The considered methods used in hypothesis assessment, which belong to a broader class of block-matching algorithms, are modified sum of absolute differences (MSAD), normalized cross correlation (NCC), census transform (CT) and modified census transform (MCT). The methods are compared w.r.t. their qualitative results for depth estimation and are presented for artificial and real-world data.

We present a new robust approach to the detection of rail surface disruptions in high-resolution images by means of 21/2D image analysis. The detection results are used to determine the condition of rails as a precaution to avoid breaks and further damage. Images of rails are taken with color line scan cameras at high resolution of about 0.2 millimeters under specific illumination to enable 21/2D image analysis. Pixel locations fulfilling the anti-correlation property between two color channels are detected and integrated over regions of general background deviations using so called cross-channel co-occurrence matrices, a novel variant of co-occurrence matrices introduced as part of this work. Consequently, the detection of rail surface disruptions is achieved with high precision, whereas the unintentional elimination of valid detections in the course of false and irrelevant detection removal is reduced. In this regard, the new approach is more robust than previous methods.

For line scan imaging the resolution along the direction of movement can be improved. Knowledge on the translational motion and imaging process enables the derivation of efficient super-resolution processing along the direction of movement. Examples of improved image resolution for partially overlapping image lines with known displacements and estimated blurring are given.

Non-negativematrixfactorizationofaninputdatamatrixintoamatrixofbasisvectorsandamatrixofencodingcoefÞcientsisasubspacerepresentationmethodthathasattractedattentionofresearchesinpatternrecognitionintherecentperiod. Wehave explored crucial aspects of NMF on massive recognition experiments with the ORL database of faces which includeintuitively clear parts constituting the whole. Using a principal changing of the learning stage structure and by formulatingNMF problems for each of a priori given parts separately, we developed a novel modular NMF algorithm. Although thisalgorithm provides uniquely separated basis vectors which code individual face parts in accordance with the parts-basedprinciple of the NMF methodology applied to object recognition problems, any signiÞcant improvement of recognitionrates for occluded parts, predicted in several papers, was not reached. We claim that using the parts-based concept in NMFas a basis for solving recognition problems with occluded objects has not been justiÞed.Keywords: non-negative matrix factorization, subspace data representation, image object recognition

In recent years, nonnegative matrix factorization (NMF) methods of a reduced image data representation attracted the attention of computer vision community. These methods are considered as a convenient part-based representation of image data for recognition tasks with occluded objects. A novel modification in NMF recognition tasks is proposed which utilizes the matrix sparseness control introduced by Hoyer. We have analyzed the influence of sparseness on recognition rates (RRs) for various dimensions of subspaces generated for two image databases, ORL face database, and USPS handwritten digit database. We have studied the behavior of four types of distances between a projected unknown image object and feature vectors in NMF subspaces generated for training data. One of these metrics also is a novelty we proposed. In the recognition phase, partial occlusions in the test images have been modeled by putting two randomly large, randomly positioned black rectangles into each test image.

In recent years non-negative factorization (NMF) methods of a reduced image data representation attracted the attention of computer vision community. These methods are considered as a convenient part-based representation of image data for recognition tasks with occluded objects. In the paper two novel modifications of the NMF are proposed which utilize the matrix sparseness control used by Hoyer. We have analyzed the influence of sparseness on recognition rates (RR) for various dimensions of subspaces generated for two image databases. We have studied the behaviour of four types of distances between a projected unknown image object and feature vectors in NMF-subspaces generated for training data. For occluded ORL face data, Euclidean and diffusion distances perform better than Riemannian ones, not following the overall expactation that Euclidean metric is suitable only for orthogonal basis vectors. In the case of occluded USPS digit data, the RR obtained for the modified NMF algorithm show in comparison to the conventional NMF algorithms very close values for all four distances over all dimensions and sparseness constraints. In this case Riemannian distances provide higher RR than Euclidean and diffusion ones. The proposed modified NMF method has a relevant computational benefit, since it does not require calculation of feature vectors which are explicitly generated in the NMF optimization process.

We present a multi-line-scan light-field image acquisition and processing system designed for 2.5/3-D inspection of fine surface structures in industrial environments. The acquired three-dimensional light field is composed of multiple observations of an object viewed from different angles. The acquisition system consists of an area-scan camera that allows for a small number of sensor lines to be extracted at high frame rates, and a mechanism for transporting an inspected object at a constant speed and direction. During acquisition, an object is moved orthogonally to the camera’s optical axis as well as the orientation of the sensor lines and a predefined subset of lines is read out from the sensor at each time step. This allows for the construction of so-called epipolar plane images (EPIs) and subsequent EPI-based depth estimation. We compare several approaches based on testing a set of slope hypotheses in the EPI domain. Hypotheses are derived from block matching, namely the sum of absolute differences, modified sum of absolute differences, normalized cross correlation, census transform, and modified census transform. Results for depth estimation and all-in-focus image generation are presented for synthetic and real data.

In this paper, we present a novel method for multi-frame superresolution (SR). Our main goal is to improve the spatial resolution of a multi-line scan camera for an industrial inspection task. High resolution output images are reconstructed using our proposed SR algorithm for multi-channel data, which is based on the trainable reaction-diffusion model. As this is a supervised learning approach, we simulate ground truth data for a real imaging scenario. We show that learning a regularizer for the SR problem improves the reconstruction results compared to an iterative reconstruction algorithm using TV or TGV regularization. We test the learned regularizer, trained on simulated data, on images acquired with the real camera setup and achieve excellent results.