This PDF file contains the front matter associated with SPIE Proceedings Volume 11172 including the Title Page, Copyright information, Table of Contents, Introduction, and Conference Committee listing.

This paper aims at describing recent advances in coherent imaging based on digital holography; especially we will focus on the use of such an approach to investigate traveling acoustic waves in mechanical structures and to study mechanical properties of bio-materials.

The paper discusses the light efficiency and SNR of light field imaging systems in comparison to classical 2D imaging. In order to achieve this goal, I define the “equivalent f-number” as a concept of capturing the light gathering ability of the light field system. Since the f-number, in classical imaging systems, is conceptually also linked with the depth-of-field, I discuss an appropriate depth-of-field interpretation for light field systems.

In clinical environments and pharmaceutical industrial productions, early detection of contamination by microorganisms is a key point in terms of quality control. Determination of such contaminations relies on cultures in Petri dishes, the observation of which, through the detection of colonies of microorganisms, leads to methods enabling their determinations. However, these methods show limits in terms of speed and are rather tedious. To overcome these shortcomings, a method based on image analysis and deep learning is proposed to improve both detection of microorganism colonies in Petri dishes and quality of the quantitative determination of the contamination levels.

In the recent past years, innumerable techniques more complex than the others have emerged in computer vision. They have been applied to many fields and, thanks to the tremendous computational power one has access to nowadays, have made possible more and more elaborate applications. In this article, we propose a classification tool, using hand-crafted interpretable (statistical and digital imaging) features, in order to confirm or invalidate the presence of passengers on skilift vehicles in moutain ranges. More precisely, Linear Discriminant Analysis, which is a dimensionality reduction alongside classification technique, and its less restrictive variant Quadratic Discriminant Analysis, are applied. One of the paper’s objectives consists in illustrating the famous law of parsimony, also known as Occam’s razor, in the sense that the simpler solutions should be considered first and more complex models should be built afterwards if needed.

In classification tasks, the accuracy of classifiers depends on training data. It is known that inter-class imbalanced data degrade the classification accuracy. Previous approaches tend to use data augmentation to solve inter-class imbalance, but the possibility of intra-class imbalance has been ignored. In this paper, we propose a novel method to solve the intra-class imbalance with Generative Adversarial Networks (GAN). The key idea is to examine the distribution of training data in latent space. We experimentally demonstrate that the proposed method generates diverse images and improves classification accuracy on the CIFAR-10 dataset.

Image registration finds its applications in various fields like environmental monitoring, change detection, automatic inspection, medical imaging and remote sensing. Industrial real time systems need a fast and accurate registration to satisfy real time constraints which are very challenging. In this work a new rigid monomodal registration method is proposed, the method exploits the symmetrical property of objects to speed up registration and to achieve robustness against changing edges and less structural information. Tests on images of a large semi-opaque containers data set show its effectiveness compared to classical methods.

In order to ease visual inspections of exterior aircraft fuselage, new technical approaches have been recently deployed. Automated UAVs are now acquiring high quality images of the aircraft in order to perform offline analysis. At first, some acquisitions are annotated by human operators in order to provide a large dataset required to train machine learning methods, especially for critical defects detection. An intrinsic problem of this dataset is its extreme imbalance (i.e there is an unequal distribution between classes): the rarest and most valuable samples represent few elements among thousands of annotated objects. Deep Learning-only based approaches have proven to be very effective when a sufficient amount of data is available for each desired class, whereas less complex systems such as Support Vector Machine theoretically need less data, and few-shot learning dedicated methods (Matching Network, Prototypical Network, etc.) can learn from only few examples. Those approaches are compared on our applicative case. Preliminary results show the existence of empirical frontiers in term of training dataset volume that indicate which approach might be favored. Based on those results, we propose a method to combine different approaches in order to achieve best performances on categorical accuracy, with special attention paid to underrepresented category.

The purpose of this paper is to explore the use of Fully Convolutional Neural Networks (FCN) to perform a semantic segmentation of deflectometric recordings for quality control of reflective surfaces. The proposed method relies on a U-Net network to identify the location and boundaries of the object, and the possible defective areas present, by performing a pixel-wise classification based on local curvatures and data modulation. Experiments performed on a real industrial problem demonstrate that the combination of geometric and photometric information enables the identification of a wider variety of shape and texture imperfections, with predictions closely correlated with the visual impact of the defects. The research also highlights the relevance of the labeling process and human inspection limits, and suggestions are presented for a near-term industrial utilization.

Reflectance Transformation Imaging (RTI) is a multi-light-based imaging technique that can provide relevant information on both local micro-geometry and visual appearance of a studied surface. The local angular reflectance is modelled to allow the relighting of the surface appearance under any arbitrary light direction. The methods used to model the local reflectance of each pixel are mainly PTM (2nd order polynomial functions), HSH (Hemispherical Harmonics) and more recently DMD (Dissrete Modal Decomposition). For all these methods, a uniform distribution of the light positions over the hemisphere is an implicit hypothesis. However, it’s impossible to satisfy this condition in practice. As a result of this non-homogeneous distribution, several artifacts can affect the reconstruction and alter the quality of the visual appearance assessment. To address this issue, we proposed a methodology consisting in the estimation of the spatial distribution of the lighting directions used during RTI acquisitions, based on a local density estimation. These local density values are then used to weight the Least Squares regression, and thus to correct the contributions of each image of the RTI acquisitions. This methodology is applied on two metallic surfaces with visual singularities. From presented results, it can be concluded that it is necessary to take into account this non-uniformity in order not to alter the quality of RTI data and subsequent inspection tasks.

In Japan, the age-related deterioration of many public roads, which were constructed in the 1960s, is demanding maintenance solutions. We propose a convolutional neural network (CNN)-based method to convert the original image to a pseudo normal road surface image. The converter selectively replaces only the abnormal data of an image with pixels corresponding to normal features, thereby creating an output image without abnormal parts. We aim to detect anomalies on the road surface by calculating the difference between a raw input image and the generated PNI image. Our experimental results confirm the effectiveness and usefulness of our method.

Multispectral measurement configurations are proposed as alternative tools for on-line moisture assessment for materials or parts on conveyers. In a first imaging configuration combining high speed Short Wave Infrared Range (SWIR) linear camera and lasers used as backlighting sources, images are acquired in several spectral bands and further processed to enhance information related to moisture level of pharmaceutical capsules. A second scanner architecture, based on a rotating polygonal mirror and an infrared multispectral fibered source operated in the SWIR spectral range, is also proposed to map water content of diffuse materials. Qualitative imaging and measurements capability of both configurations have been investigated in laboratory or in representative sorting machine environment.

Aramid composites very often occur in the form of laminates, resulting from the combination of layers of aramid fabric joined by various resin types. The most effective methods in the testing of composites reinforced with fibres are active methods of infrared thermography. In non-destructive thermographic testing, defect detection is carried out by means of image (thermogram) analysis. During tests, a sequence of thermograms consisting mainly of several hundred or more individual images is recorded. An effective analysis of such a large number of images without the use of appropriate algorithms is practically impossible. In order to improve the visual representation of defect occurrence in the thermograms, special image processing methods are utilised, such as Principal Component Analysis. . Selected results from these analyses are presented.

The detection of cracks on metallic parts is critical in many industrial applications. Thermography is one of the most effective methods for this purpose. In laser heating the infrared images are strongly affected by laser reflection or secondary emission. To avoid this problem, during the measurement, the laser beam can be periodically alternated between on and off states. For this technique to be effective, it is important to take into account the phase of the laser (e.g. on/off state) at each image frame. Due to the noisy and non-stationary nature of the observed real and apparent temperature of the surface at the location of the laser spot, it is challenging to detect the phase of laser pulse. In this work, Bayesian recursive filtering (BRF) is successfully applied for achieving this goal. This method can seamlessly account for non-periodic conditions and inaccuracies in laser modulation frequency and sampling rate.

Tomographic diffractive microscopy (TDM) is an imaging technique, which allows for recording the refractive index of unlabelled transparent specimens. Based on diffraction theory, it can be implemented in transmission or in reflection. In this paper, a new TDM data acquisition and reconstruction method is proposed. The purpose is to use the mirror effect of a reflecting material to establish a double illumination system. Neglecting backward-diffracted fields, the setup reduces to a double-transmission TDM, which combined with an azimuthal rotation of the illumination, allows for faster and simplified acquisitions. We also point out a new demodulation method based only on Fourier transforms.

Various defects can appear on the surface of a potato, producing an adverse effect on their market price. For several years, manual methods have been applied to classify this tuber, which caused certain drawbacks such as: high-cost, high-processing time and subjective results. In this paper we introduce a deep-learning based method to classify and localize defects in potatoes with the aim to automate the quality control task. An extensive dataset was created including six potato categories: healthy, damaged, greening, black dot, common scab and black scurf. Then, a convolutional neural network (CNN) was trained with this dataset in order to achieve the classification task. We also propose to leverage the localization capability of the trained network to localize the region of the classified defect. Finally, a global evaluation was done in a test set, where 4 different sides images were taken into account to represent one tuber. Experimental results with different CNN architectures are shown. We achieved an average F1-score of 0.94 for the classification task. The localization performance is measured qualitatively by a heat map output, which shows that the proposed method accurately localize the defects.

In this paper, an effective and novel automatic learning solution for the quality control of metallic objects surfaces is proposed, which can be seamlessly integrated into the industrial process. Such a system requires a coaxial illuminator to capture the object view with a single camera while lighting it with structured light: in this way, the object surface can be viewed in time as a dynamic scene under different illumination conditions. By relying on a linear model to describe the expected evolution of the light over the object of interest, the Residuals of Linear Evolution of Light (RLEL) algorithm is derived with the specific aim of identifying and characterizing anomalies and defects through the residuals of a least square approach. Then, a novel learning strategy is developed that exploits the model-based RLEL descriptor and thus promotes itself as an alternative strategy to the black box approach of Convolutional Neural Networks (CNNs). By combining both the data-driven and the model-based learning approaches to perform the inspection task, an Hybrid Learning (HL) procedure is defined: in a first phase, the HL exploits an Encoder-Decoder network to incorporate the model-based description while, in a second phase, it uses only the pre-trained encoder to drive the learning process of a 3D-CNN. In doing so, the proposed procedure reaches interesting results that exceed also the performance of state-of-the-art 3D-Inception and 3D-Residual networks.

In this paper a method is proposed for estimating the orientation of industrial parts. Classical 2d images of the part are used to train a deep neural network and infer the part pose by computing its quaternion representation. Another innovative point of this work is in the use of synthetic data which are generated on the fly during the network training from a textured cad model placed in a virtual scene. This way overcomes the difficulties to obtain pose ground truth images from real images. At the same time, using the cad model with several lighting conditions and material reflectances allows to anticipate challenging industrial situations. As a first step of the method, the part is separated from the background using a semantic segmentation network. Then a depth image of the part is produced employing an encoder-decoder network with skip connections. Finally, the depth map is associated with the local pixels coordinates to estimate the part orientation with a fully connected network using a SO(3) metric loss function. This method can estimate the part pose from real images with visually convincing results suitable for any pose refinement processes.

With the development of green energy and its means of production, more and more companies chose to build solar panel farms. However, those technologies remain relatively expensive to maintain, and prone to damages (due to natural hazards, or internal defects). Since any kind of damage on a panel cell drastically reduce a panel's efficiency, solar panels must be kept under tight supervision. With more solar panel that must be checked for damage relatively often, a cheap, accurate and fast way to find those damages must be settled. Some processes have been developed to identify panels in a true color image [1], and various ways to identify defective panels exist through image processing [2], [3] or other ways [4]. On another hand, handmade features suggest the input data obeys to some specific conditions (color, illumination), and small changes can impact accuracy. CNN [5], however, can be trained to face such changes with the appropriate dataset, and therefore be more resilient . They represent a reliable solution for identification and classification of complex features [2], [6], and can be improved more easily than handmade feature detection. In this paper is detailed the pipeline of such process, combining the straightforward approach of handmade feature detection for preprocessing to reduce the input’s complexity, with the resilience of neural networks for the final identification. Detailed explanations for the different steps of the process are given: Dataset acquisition, preprocessing, and finally classification. The various leads that were followed to improve the quality of the results are also given, before comparing results with a previously used handmade detection process, and finally proposing a web user interface to exploit this process, and enrich its dataset.

Convolutional neural networks have achieved excellent results in pedestrian detection state of the art. They are able to learn features from raw images which makes them easy, practical and robust for multiple visual classification tasks. In this paper, we propose a further improvement of convolutional neural networks using saliency detection. First, we use contourlet transform for saliency detection to generate a region of interest (ROI). The generated saliency maps are then used to feed the convolutional network which will be used for both feature extraction and classification. The paper contribution is two fold : (1) We use saliency detection as a filter to remove the noisy information in the background, which allow the network to converge faster during the training process. (2) Saliency reduced complexity of the road scene which improve significantly the CNN classification performance. Experiments conducted on INRIA and Pascal VOC datasets achieves state-of-the-art performance.

In capital goods industry, there are some components that are employed for safety purposes and, due to this fact, parts are subjected to high quality control demands. This is especially relevant for the case safety components that contain welds because of the inherent complex process and the likelihood of defects appearance. In this context, this work presents a machine vision system that was employed for replacing costly quality control procedures based on visual inspection. This was possible thanks to the proper design of all the machine vision system components including the image processing algorithm. As a special feature of the system, it has to be highlighted the low cycle time of the production process (<2s), which stablished some requirements on the image processing algorithms. During the inspection system development, the main efforts were concentrated for obtaining a reliable and balanced database of defective and non-defective parts images useful to train the classification model. At this respect, the main contributions consisted of image analysis software development and visual curation of data. As a result, tailor made filters were developed that allowed together with color information the identification of common flaws, as Lacks of Fusion (LoF). Due to the high amount of inspected samples, a preliminary deep learning based model was developed that included these filters with the aim of increasing defect detection accuracy.

The use of photographic films is not totally obsolete, photographers continue to use this technology for quality in terms of aesthetic rendering. A crucial step with films is the digitization step. During the scanning process, dust, scratch and hair (artefacts) are a real problem and greatly affect the quality of final images. The artefacts correction has become a challenge in order to preserve the quality of these photos. In this article, we present a new method based on deep learning with an encoder-decoder architecture to detect and eliminate artefacts. In addition, a dataset has been created to carry out the experiments.

This paper proposes a fully automated vision system to inspect the whole surface of crankshafts, based on the magnetic particle testing technique. Multiple cameras are needed to ensure the inspection of the whole surface of the crankshaft in real-time. Due to the very textured surface of crankshafts and the variability in defect shapes and types, defect detection methods based on deep learning algorithms, more precisely convolutional neural networks (CNNs), become a more efficient solution than traditional methods. This paper reviews the various approaches of defect detection with CNNs, and presents the advantages and weaknesses of each approach for real-time defect detection on crankshafts. It is important to note that the proposed visual inspection system only replaces the manual inspection of crankshafts conducted by operators at the end of the magnetic particle testing procedure.

Designing the optimal architecture of neural networks is an important issue. However, since this is difficult even for experienced experts, automatic optimization of the network architecture is required. In this study, we regard this issue as a combinatorial optimization problem, and utilize genetic algorithm to optimize the network architecture. Because training the networks, which are represented by individuals in GA, takes a long time, a novel method to reduce the training time by inheriting the weights of the trained network is proposed. From experimental results, our proposed method achieved the time reduction and higher accuracy than a conventional method.

In sheet metal production the quality of a cut is determined by the quality of the cut edge and is of crucial importance. One parameter affecting the quality of the cut edge surface is its roughness. In order to determine the roughness, the depth information is required. The common methods for acquiring depth information are very time consuming and therefore not suitable for a quick roughness evaluation. We present a method for a quick roughness evaluation by means of 2D image processing. It is shown that, given a proper dataset, a convolutional neural network can be trained to identify image features that correlate highly with the roughness of the edge surface and learn how to weight these features correctly. This makes the neural network capable of providing a quick and accurate statement about the roughness of the edge surface based on an image.

In dermatology, image processing allows non-contact and non-invasive metrological measurements. Psoriasis is an incurable skin disease with an unknown origin. One of the most important tasks in the treatment of psoriasis is to evaluate the degree of the illness following a severity score. Dermatologists use visual and tactile senses to assess the lesions severity. In this article, we propose an automated methodology for assessing objectively the severity of psoriasis by measuring the physical parameters of the skin. Thus, from the colorimetry and geometry obtained by photometric-stereo, we determine the level of erythema and skin thickness. Our results show that for a low acquisition time, the scores obtained are highly correlated with those of dermatologists.

In this research two methods for recognizing water droplets that are formed during dropwise condensation on the flat and pillared substrates are presented. The aim of these methods is to binarize the gray scale images of the droplets taken by a CCD camera in order to extract the information related to the droplets size and density.

In recent years, driven by new standards and brand image, product quality has become highly critical in the beverage industry. Product quality does not concern the beverage in and of itself only, but also the extrinsic aspects, particularly the presence of very small foreign bodies. Without reliable and affordable methods for detecting such foreign bodies in semiopaque bottles, many manufacturers still opt for human visual inspection on the production line. Optical advanced technological systems such as X-ray based systems started to be used but such technologies are expensive and intrusive. Also the visible optical domain offers an alternative and an advantageous solution. Thus a novel detection method based on two robust criteria such as kurtosis and connected components is proposed to give a further validation of the detection output. The method applies on static images acquired by a RGB camera and the experimental results show its effectiveness in detecting random foreign objects regardless their size or their transparency in various semi opaque bottles.

We propose a defect detection method of solders on a printed circuit board using X-ray CT inspection system and Adversarial Autoencoder (AAE)^[1] . We obtain sliced images of the solder using X-ray CT and extract their features that follow the standard normal distribution by using AAE. Then, the solder defects are detected by Hotelling's T square^[2]. As a result of experiments, we show that we can classify normal and anomalous data samples completely on the condition of training with large normal samples and small anomalous samples.

In this paper, we address the problem of automatic robotic inspection in two parts: first, automatic selection of informative viewpoints before the inspection process is started, and, second, automatic treatment of the acquired 3D point cloud from said viewpoints. We apply our system to detecting defects on aircraft Electrical Wiring Interconnection System (EWIS) in order to comply with the growing amount of safety regulations such as interference and allowable bend radius of cables in mechanical assemblies.

Widely used for surface slopes measurements and for three-dimensional shape reconstruction, deflectometry is a particularly powerful technique that can also be applied for defects detection on specular surfaces. In the visible domain, deflectometry is usually based on the projection of complex encoded light patterns and necessitates heavy processing that makes it not suitable for inline inspection. In this paper, A new deflectometry based approach that is more adapted for inline inspection of linearly moving parts (parts on conveyors) is proposed. Based on a more affordable and a simpler hardware setup, the new approach allows at the same time for a proper localization and a precise geometrical quantification of any defects on the scanned specular surfaces. The proposed approach uses a fast and simple processing algorithm that lends itself very well to real-time inspection. The new method is tested and validated in laboratory for the inspection of defects on specular surfaces of plastic parts.

During the last two decades the number of visual odometry algorithms has grown rapidly. While it is straightforward to obtain a qualitative result, if the shape of the trajectory is in accordance with the movement of the camera, a quantitative evaluation is needed to evaluate the performances and to compare algorithms. In order to do so, one needs to establish a ground truth either for the overall trajectory or for each camera pose. To this end several datasets have been created. We propose a review of the datasets created over the last decade. We compare them in terms of acquisition settings, environment, type of motion and the ground truth they provide. The purpose is to allow researchers to rapidly identifies the datasets that best fit their work. While the datasets cover a variety of techniques to establish a ground truth, we provide also the reader with techniques to create one that were not present among the reviewed datasets.

This paper proposes a novel approach that performs extrinsic parameter estimation of a camera installed in a man-made environment using a single image. The problem of extrinsic parameter calibration is identical to 6DoF (six-degrees of freedom) localization problem of the camera. We take advantage of line information that is usually present in the man-made environment such as inside of the building. Our approach only requires a flat surface map for a 3D environment model which can be easily obtained from the blueprint of the artificial environment (e.g., CAD data). In order to manage the complicated 6DoF search problem, we propose a novel image descriptor defined in quantized Hough space to perform 3D-2D matching process between line features from the 3D flat surface model and the 2D single image. The proposed method can robustly estimate the complete extrinsic parameters of the camera, as we demonstrate experimentally.

Wisely exploiting scattered radiation, considered as noise until now, for image purpose is one of relevant techniques for emerging tomographic imaging. In this work we propose a new concept of Compton Scattering Tomography (CST) based on a fixed source and a ring of detectors passing through the source. The modelling of this new CST modality leads to a new Radon transform (RT) on a special family of circular arcs. The inversion of this RT is calculated. Simulation results illustrate the viability and performances of the proposed system. It has the advantage of being able to scan both small and large objects, which makes it easy to apply to several fields as quality control in non destructive evaluation, biomedical and cultural heritage object imaging.

This study investigates the combination of Mueller imaging polarimetry with machine learning for the automated optical classification of raw materials. It shows that standard image classification techniques based on support vector machines or deep neural networks can readily be applied to polarimetric data extracted from Mueller matrix measurements. The feasability of such an approach is empirically demonstrated through the classification of multispectral depolarization images of real-world materials (banana, wood and foam samples).

Some historical musical instruments are still played today, and are prone to be affected by mechanical wear of superficial varnishes in direct contact with the musicians. In a previous work, an ad-hoc monitoring plan for historical violins, that involves the use multiple non-invasive analytic techniques, achieved good performance. However, the high number of techniques is a limitation if we want to perform frequent checks. In this work, our aim is to rely only on UV induced fluorescence images for performing a fast, preliminary check and then, if a possible alteration is detected, conduct spectroscopic analyses, which are more precise but also more time consuming. In this study, we explore the a-contrario framework in order to allow for the automatic detection of significant changes in the superficial varnishes. The difficulty of detecting the changes is represented by the need to define the significance of a change, in the absence of a ground truth provided by the expert. Tests performed on samples that simulate the effect of surface alteration during time show promising results.

The development of a spectral difference-based statistical processing of hyperspectral images is provided in this article. Kullback-Leibler pseudo-divergence function, which was specifically developed for the metrological processing of hyperspectral images, is used at the foundation of the statistics. As a demonstration of its use, the proposed statistics are used in visualising surface variability within a set of pigment patches. It is then further exploited to detect anomalies and deterioration that occur on the patches.

More and more electronic products use non-smooth surfaces to improve user experience. Computer vision based quality control is very important for these products. Since images taken from the non-smooth surfaces have complex micro textures and low contrasts, it is very challenging to detect defects in the images. To solve this problem, this paper proposes a training-free method which first enhances defects and then detect them accurately. At the phase of defect enhancement, pixel-level saliency is first calculated by two novel features named localglobal intensity difference and local intensity aggregation, and then an iterative enhancement approach named accumulated aggregation shifting (AAS) is proposed to shift each pixel’s intensity according to its saliency. At the phase of defect detection, two statistic models, including linear distribution or exponential distribution, are fitted by the shifting results of AAS at different iterations. Based on the fitted statistic models, defective pixels are defect-free pixels are accurately classified by risk minimization. Experimental results prove that the proposed approach is effective in detecting defects on non-smooth surfaces of real industrial products.

Visual inspection of surfaces is a complex sensory process that depends on several factors related to the observer [1]–[3] (his knowledge, experience, tiredness or the purpose of observation), the object observed and the light environment. In order to reduce the subjectivity and to improve the robustness of the analysis of the appearance quality of the inspected surfaces, multi-lighting imaging techniques like RTI (Reflectance Transformation Imaging) are being used more and more. However, and to make the most of this technique, it is necessary to control the acquisition and processing, especially for highly reflective surfaces and / or having particular surface topographies. Thus, we propose an innovative implementation based on the RTI technique by increasing it with an acquisition and a processing in high dynamic range (HDR), adapted and adaptive. This coupling makes it possible to avoid the low dynamic range (LDR) of the cameras and therefore ensures a faithful estimate, for each pixel, of the angular component of the reflectance of the surface inspected, in particular in the case of non-Lambertian surfaces, and heterogeneous surfaces in terms of material and / or geometric texturing. Beyond appearance reconstruction, maps associated with geometric and statistical indicator maps (local slope/kurtosis/skewness/entropy) are used to show the relevance and performance of the proposed HDR-RTI methodology.

In this paper, a novel system that operates home appliances at arbitrary positions in a room is proposed based on a “command space” associated with the operation of the home appliance. In the proposed system, a hand waving gesture is used to operate the home appliances. First, three-dimensional (3D) positions are extracted from the hand waving gestures at two different positions, and a command space is set up based on the extracted 3D positions. Thus, it is possible to operate the home appliances freely in an arbitrary place chosen by the user by installing the user-definable command space. Next, detailed operations are performed by hand waving in the command space. Experiments were conducted to confirm that detailed operations, such as TV channel switching, can be executed from different places using the proposed system.

This paper presents a passive polarimetry method using a division of aperture optical device in order to segment the weld pool surface of the welding. Due to the high specularity of the surface of the weld pool, we use polarimetric data in order to facilitate it’s segmentation in the liquid state. In this paper, we propose to combine two segmentation methods: Watershed Transform and the Level Set to ensure a better result. Our ultimate goal of this study is to provide real-time quality control of the surface of molten metal during the welding process while utilizing the additional information provided by the polarimetric data.

A Polarization Filter Array (PFA) camera is an imaging device capable of analyzing the polarization state of light in a snapshot way. These cameras exhibit spatial variations, i.e. nonuniformity, in their response due to optical imperfections introduced during the nanofabrication process. Calibration is done by computational imaging algorithms to correct the data for radiometric and polarimetric errors. In this paper, we review existing calibration procedures, and show a practical implementation result of one of these methods applied to a commercially available PFA camera.

Quality control is a key requirement of the Fourth Industrial Revolution, especially in manufacturing. For high value technical parts, geometrical and appearance quality need to be guaranteed. The quality control of the part, directly on the manufacturing line, can be achieved with non-contact sensors. With a robust quality sensor, the manufacturing process could be adjusted in real time to always achieve the optimal quality. For quality control, visible spectrum imaging is commonly used. However, more information could be obtained from broader spectrum imaging and also from the degree of linear polarization of the light from the part. We have developed a simple polarimetric imaging system to verify whether linear polarization is capable of amplifying geometric and appearance defects. We evaluate the use of this polarimetric imager given the industrial constraints of injection molding. We compare supervised classification performances on an injection molded parts dataset, using polarimetry and non-polarized images. We compare the performances of different machine learning pipelines with hand-crafted feature extraction and Deep Transfer Learning. With its industrial robustness, polarimetry could be a valuable addition to non-contact imagers for geometric and appearance quality control.

Object detection in road scenes is an unavoidable task to develop autonomous vehicles and driving assistance systems. Deep neural networks have shown great performances using conventional imaging in ideal cases but they fail to properly detect objects in case of unstable scenes such as high reflections, occluded objects or small objects. Next to that, Polarized imaging, characterizing the light wave, can describe an object not only by its shape or color but also by its reflection properties. That feature is a reliable indicator of the physical nature of the object even under poor illumination or strong reflections. In this paper, we show how polarimetric images, combined with deep neural networks, contribute to enhance object detection in road scenes. Experimental results illustrate the effectiveness of the proposed framework at the end of this paper.

Facial expressions and hand gestures are recognized as a part of human emotions especially in a feeling of fatigue signs. In computer vision research, a positioning of hand over face is one of the challenging problem caused by difficulty of the difference of skin color for hands and face. In this paper, we present a method for classifying six positions of the hand over face which is able to identify the signs of feeling fatigue for the visual display terminal (VDT) workers. We apply a deep learning method in order to compare with the methods used the face and skin colors detection, processing the edge detection and feature extraction algorithms. In addition, GoogleNet is used for training a data set made in the simulated VDT workers environment. The data set includes 1,440 images, the participants from several countries, Egypt, Japan, Bangladesh, Mongolia and Rwanda to cover a wide range of skin tones. The data set is categorized by six groups of the positions of hand over face. These groups consist of hands on a forehead, eyes, nose, mouth, right and left face. The experiments were performed using MATLAB to implement our proposed method. The system achieved average recognition ratio 99.3 % in all hand over face gestures.

Visual tracking gives a trajectory of a person’s movement in a video. It is an important element of human’s behavior analysis. However, the performance of the most visual tracking methods does not satisfy precision and speed for use in the real world. High accuracy and low computation complexity are demanded but these two have a relation of trade-off. Therefore, we aim to propose a tracking method that balances accuracy and computational complexity. In this paper, we proposed a method named Dual Cost Graph (DCG)-Tracker using the graphs referred to as a clique graph and a flow network. We evaluated DCG-Tracker with PNNL parking lot 1 dataset^{1, 2} and showed that it balances between accuracy and speed.

A 3D reality world used for medical therapies needs clear images without defective and noisy pixels. This is essential for portative equipments used by several patients inside the same building. To reduce the resources of the system, a corrective and denoising implementation has been implemented based on FPGA. Two low resources algorithms are proposed for defective pixel correction and noise reduction. The correction algorithms are integrated according to ”Bayer Color Filter Array” for stereoscopic video system respecting real-time constraints. State-of-the-art denoising algorithm shows that implementations use too many resources for the target system, therefore a new low resource denoising algorithm is presented. It is derived from the efficient Non-Local Means (NLM) denoising algorithm which is hardly implementable in a FPGA system. The new denoising method discussed in this paper does not use any multiplier and only consumes 3.8k ALUTs.

A number of high-quality depth imaged-based rendering (DIBR) pipelines have been developed to reconstruct a 3D scene from several images taken from known camera viewpoints. Due to the specific limitations of each technique, their output is prone to artifacts. Therefore, the quality cannot be ensured. To improve the quality of the most critical and challenging image areas, an exhaustive comparison is required. In this paper, we consider three questions of benchmarking the quality performance of eight DIBR techniques on light fields: First, how does the density of original input views affect the quality of the rendered novel views? Second, how does disparity range between adjacent input views impact the quality? Third, how does each technique behave for different object properties? We compared and evaluated the results visually as well as quantitatively (PSNR, SSIM, AD, and VDP2). The results show some techniques outperform others in different disparity ranges. The results also indicate using more views not necessarily results in visually higher quality for all critical image areas. Finally, we have shown a comparison for different scene’s complexity such as non-Lambertian objects.

Recently, visual object tracking has become a vital role in many different applications. It spans various applications like robotics, video surveillance, and intelligent vehicle navigation. In the Counter-UAV (Unmanned Aerial Vehicle) system, in order to shoot down illegal UAVs flying in the sky, it is very important task to accurately track UAV by an image sensor. However, accurate target tracking is a difficult problem because the distance to the target is far, there are severe shape changes, and the target is often obscured. The remarkable method in object tracking is correlation filter based tracker like KCF (Kernelized Correlation Filter). Although the tracker is very competitive in terms of computation time and performance among the state-of-the-art trackers, it has problems related to target drift in tracking process. It comes from that target objects undergo significant appearance variation due to deformation, partial occlusion, camera shaking, and illumination changes. To mitigate the drift problem of the target in CUAV application, we propose a tracking method which uses KCF and adaptive thresholding. When the correlation response is less than a certain value during KCF process, the result of adaptive thresholding adjusts the tracking point. In the experiment, the proposed method was verified by using image sequence obtained by a monochrome camera on the CUAV system.

Thermal imaging is nowadays widely used in many industrial areas for non destructive analysis. This paper focuses on the detection by active thermography of cracks in a transparent and specular piece of glass. A continuous laser scans the target by creating a hot spot that is acquired by a thermal camera. A thermal map is computed by averaging of thermal measurements for all the positions of the hot spot on the target. Then edge filters are applied to highlight the cracks. Experimental results demonstrate the validity of the material damage estimation in a short processing time.

Texture discrimination was studied a lot for texture classification/recognition in image databases, but less under the metrological point of view. In this work, we focused on the metrological behaviour related to the human vision for Control Quality purposes. Inside this study, we introduce as a pair a novel texture feature associated to an adapted similarity measure. The main idea was to define a compact representation adapted from the human visual characteristics in order to obtain an accurate description of the texture. Combined to an adapted similarity measure, the obtained pair feature/similarity becomes highly efficient. Performance Classification of the proposed texture feature is assessed on six popular and challenging databases used to provide the reference results in the state-of-the-art. Obtained results show the efficiency and the robustness of the proposed pair feature/similarity measure defined by the relocated Colour Contrast Occurrence Matrix.

Automated fiber placement (AFP) is an advanced manufacturing technology that has led to an increased rate of production in composite materials. At the same time, the need for adaptable and fast inline control methods increases. Existing inspection systems make use of handcrafted filter chains and feature detectors tuned for a specific measurement method by domain experts. These methods hardly scale to new defects or different measurement devices. In this paper, we propose to formulate AFP defect detection as an image segmentation problem that can be solved in an end-to-end fashion using artificially generated training data. We employ a probabilistic graphical model to generate training images and annotations. We then train a deep neural network using a recent architecture designed for image segmentation. This leads to an appealing method that scales well with new defect types and measurement devices and requires little real world data for training.

In recent years, driven by new standards and brand image, performant quality control has become highly important in the beverage industry. In the absence of reliable and affordable methods for detecting foreign bodies in semi-opaque bottles, many manufacturers still opt for human visual inspection on the production line. Advanced technological systems such as X-ray based systems started to be used but such technologies are expensive and intrusive. The visible domain offers an alternative and advantageous solution. In this work a recurrent boosting method is proposed and extended to dynamic information to detect small transparent objects. The proposed method works on images video acquired by an RGB camera, it considers an ensemble classification strategies that includes temporal and dynamical features in addition to the classical spatial image features such as texture and color. The experimental results show a better performance with time series and dynamic information, the effectiveness of the method is demonstrated in detection of random foreign objects regardless their size or transparency in different semi-opaque bottles.