Digital Holography Microscopy (DHM) is a new 3D measurement technique that exists since Charge Coupled Devices (or CCD cameras) allow to record numerically high resolution images. That opens a new door to the theory of holography discovered in 1949 by Gabor: the door that masked the world of digital hologram processing. A hologram is a usual image but that contains the complex amplitude of the light coded into intensities recorded by the camera. The complex amplitude of the light can be seen as the combination of the energy information (squared amplitude modulus) with the information of the propagation angle of the light (phase of the amplitude) for each point of the image. When the hologram is digital, this dual information associated with a diffractive model of the light propagation permits to numerically investigate back and front planes to the recorded plane of the imaging system. We understand that 3D information can be recorded by a CCD camera and the acquisition rate of this volume information is only limited by the acquisition rate of the unique camera. For each digital hologram, the numerical investigation of front and back regions to the recorded plane is a tool to numerically refocus objects appearing unfocused in the original plane acquired by the CCD.<p>This thesis aims to develop general and robust algorithms that are devoted to automate the analysis process in the 3D space and in time of objects present in a volume studied by a specific imaging system that permits to record holograms. Indeed, the manual processing of a huge amount of holograms is not realistic and has to be automated by software implementing precise algorithms. In this thesis, the imaging system that records holograms is a Mach-Zehnder interferometer working in transmission and studied objects are either of biological nature (crystals, vesicles, cancer cells) or latex particles. We propose and test focus criteria, based on an identical focus metric, for both amplitude and phase objects. These criteria allow the determination of the best focus plane of an object when the numerical investigation is performed. The precision of the best focus plane is lower than the depth of field of the microscope. From this refocus theory, we develop object detection algorithms that build a synthetic image where objects are bright on a dark background. This detection map of objects is the first step to a fully automatic analysis of objects present in one hologram. The combination of the detection algorithm and the focus criteria allow the precise measurement of the 3D position of the objects, and of other relevant characteristics like the object surface in its focus plane, or its convexity or whatever. These extra relevant measures are carried out with a segmentation algorithm adapted to the studied objects of this thesis (opaque objects, and transparent objects in a uniform refractive index environment). The last algorithm investigated in this research work is the data association in time of objects from hologram to hologram in order to extract 3D trajectories by using the predictive Kalman filtering theory. <p>These algorithms are the abstract bricks of two software: DHM Object Detection and Analysis software, and Kalman Tracking software. The first software is designed for both opaque and transparent objects. The term object is not defined by one other characteristic in this work, and as a consequence, the developed algorithms are very general and can be applied on various objects studied in transmission by DHM. The tracking software is adapted to the dynamic applications of the thesis, which are flows of objects. Performance and results are exposed in a specific chapter. <p> / Doctorat en sciences appliquées / info:eu-repo/semantics/nonPublished

No description available.

Recently, there is a great interest in moving object tracking in the fields of security and surveillance. Object recognition under partial occlusion is the core of any object tracking system. This thesis presents an automatic and real-time color object-recognition system which is not only robust but also occlusion tolerant. The intended use of the system is to recognize and track external vehicles entered inside a secured area like a school campus or any army base. Statistical morphological skeleton is used to represent the visible shape of the vehicle. Simple curve matching and different feature based matching techniques are used to recognize the segmented vehicle. Features of the vehicle are extracted upon entering the secured area. The vehicle is recognized from either a digital video frame or a static digital image when needed. The recognition engine will help the design of a high performance tracking system meant for remote video surveillance.

We tackle the task of semi-supervised video object segmentation, i.e, pixel-level object classification of the images in video sequences using very limited ground truth training data of its corresponding video. Recently introduced online adaptation of convolutional neural networks for video object segmentation (OnAVOS) has achieved good results by pretraining the network, fine-tuning on the first frame and training the network at test time using its approximate prediction as newly obtained ground truth. We propose Flow Adaptive Video Object Segmentation (FAVOS) that refines the generated adaptive ground truth for online updates and utilizes temporal consistency between video frames with the help of optical flow. We validate our approach on the DAVIS Challenge and achieve rank 1 results on the DAVIS 2016 Challenge (single-object segmentation) and competitive scores on both DAVIS 2018 Semi-supervised Challenge and Interactive Challenge (multi-object segmentation). While most models tend to have increasing complexity for the challenging task of video object segmentation, FAVOS provides a simple and efficient pipeline that produces accurate predictions.

Deep learning ship detection in satellite optical imagery suffers from false positive occurrences with clouds, landmasses, and man-made objects that interfere with correctly classifying ships. A custom U-Net is implemented to challenge this issue and aims to capture more features in order to provide a more accurate class accuracy. This model is trained with two different systematic architectures: single node architecture and a parameter server variant whose workers act as a boosting mechanism. To ex-tend this effort, a refining method of offline hard example mining aims to improve the accuracy of the trained models in both the validation and target datasets however it results in over correction and a decrease in accuracy. The single node architecture results in 92% class accuracy over the validation dataset and 68% over the target dataset. This exceeds class accuracy scores in related works which reached up to 88%. A parameter server variant results in class accuracy of 86% over the validation set and 73% over the target dataset. The custom U-Net is able to achieve acceptable and high class accuracy on a subset of training data keeping training time and cost low in cloud based solutions. / Graduate

No description available.

La topologie algébrique, bien que domaine abstrait des mathématiques, apporte de nouveaux concepts pour le traitement d'images. En effet, ces tâches sont complexes et restent limitées par différents facteurs tels que la nécessité d’utiliser un paramétrage, l'influence de l'arrière-plan ou la superposition d'objets. Nous proposons ici des méthodes dérivées de la topologie algébrique qui diffèrent des méthodes classiques de traitement d'images par l’intégration d’informations locales vers des échelles globales grâce à des invariants topologiques. Une première méthode de segmentation d'images a été développée en ajoutant aux caractéristiques statistiques classiques d’autres de nature topologique calculées par homologie persistante. Une autre méthode basée sur des complexes topologiques a été développée dans le but de segmenter les objets dans des images 2D et 3D. Cette méthode segmente des objets dans des images multidimensionnelles et fournit une réponse à certains problèmes habituels en restant robuste vis à vis du bruit et de la variabilité de l'arrière-plan. Son application aux images de grande taille peut se faire en utilisant des superpixels. Nous avons également montré que l'homologie relative détecte le mouvement d’objets dans une séquence d'images qui apparaissent et disparaissent du début à la fin. Enfin, nous posons les bases d’un ensemble de méthodes d'analyse d'images basé sur la théorie des faisceaux qui permet de fusionner des données locales en un ensemble cohérent. De plus, nous proposons une seconde approche qui permet de comprendre et d'interpréter la structure d’une image en utilisant les invariants fournis par la cohomologie des faisceaux. / Algebraic topology, which is often appears as an abstract domain of mathematics, can bring new concepts in the execution of the image processing tasks. Indeed, these tasks might be complex and limited by different factors such as the need of prior parameters, the influence of the background, the superposition of objects. In this thesis, we propose methods derived from algebraic topology that differ from classical image processing methods by integrating local information at global scales through topological invariants. A first method of image segmentation was developed by adding topological characteristics calculated through persistent homology to classical statistical characteristics. Another method based on topological complexes built from pixels was developed with the purpose to segment objects in 2D and 3D images. This method allows to segment objects in multidimensional images but also to provide an answer to known issues in object segmentation remaining robust regarding the noise and the variability of the background. Our method can be extended to large scale images by using the superpixels concept. We also showed that the relative version of homology can be used effectively to detect the movement of objects in image sequences. This method can detect and follow objects that appear and disappear in a video sequence from the beginning to the end of the sequence. Finally, we lay the foundations of a set of methods of image analysis based on sheaf theory that allows the merging of local data into a coherent whole. Moreover, we propose a second approach that allows to understand and interpret scale analysis and localization by using the sheaves cohomology.

Neuro-fuzzy modeling is an efficient computing paradigm for system modeling problems. It mainly integrates two well-known approaches, neural networks and fuzzy systems, and therefore possesses advantages of them, i.e., learning capability, robustness, human-like reasoning, and high understandability. Up to now, many approaches have been proposed for neuro-fuzzy modeling. However, it still exists many problems need to be solved. We propose in this thesis two self-constructing rule generation methods, i.e., similarity-based rule generation (SRG) and similarity-and-merge-based rule generation (SMRG), and one hybrid learning algorithm (HLA) for structure identification and parameter identification, respectively, of neuro-fuzzy modeling. SRG and SMRG group the input-output training data into a set of fuzzy clusters incrementally based on similarity tests on the input and output spaces. Membership functions associated with each cluster are defined according to statistical means and deviations of the data points included in the cluster. Additionally, SMRG employs a merging mechanism to merge similar clusters dynamically. Then a zero-order or first-order TSK-type fuzzy IF-THEN rule is extracted from each cluster to form an initial fuzzy rule-base which can be directly employed for fuzzy reasoning or be further refined in the next phase of parameter identification. Compared with other methods, both our SRG and SMRG have advantages of generating fuzzy rules quickly, matching membership functions closely with the real distribution of the training data points, and avoiding the generation of the whole set of clusters from the scratch when new training data are considered. Besides, SMRG supports a more reasonable and quick mechanism for cluster merging to alleviate the problems of data-input-order bias and redundant clusters, which are encountered in SRG and other incremental clustering approaches. To refine the fuzzy rules obtained in the structure identification phase, a zero-order or first-order TSK-type fuzzy neural network is constructed accordingly in the parameter identification phase. Then, we develop a HLA composed by a recursive SVD-based least squares estimator and the gradient descent method to train the network. Our HLA has the advantage of alleviating the local minimal problem. Besides, it learns faster, consumes less memory, and produces lower approximation errors than other methods. To verify the practicability of our approaches, we apply them to the applications of function approximation and classification. For function approximation, we apply our approaches to model several nonlinear functions and real cases from measured input-output datasets. For classification, our approaches are applied to a problem of human object segmentation. A fuzzy self-clustering algorithm is used to divide the base frame of a video stream into a set of segments which are then categorized as foreground or background based on a combination of multiple criteria. Then, human objects in the base frame and the remaining frames of the video stream are precisely located by a fuzzy neural network which is constructed with the fuzzy rules previously obtained and is trained by our proposed HLA. Experimental results show that our approaches can improve the accuracy of human object identification in video streams and work well even when the human object presents no significant motion in an image sequence.

In this thesis, novel approaches for dense depth field estimation and object segmentation from mono, stereo and multiple views are presented. In the first stage, a novel graph-theoretic color segmentation algorithm is proposed, in which the popular Normalized Cuts 59H[6] segmentation algorithm is improved with some modifications on its graph structure. Segmentation is obtained by the recursive partitioning of the weighted graph. The simulation results for the comparison of the proposed segmentation scheme with some well-known segmentation methods, such as Recursive Shortest Spanning Tree 60H[3] and Mean-Shift 61H[4] and the conventional Normalized Cuts, show clear improvements over these traditional methods. The proposed region-based approach is also utilized during the dense depth map estimation step, based on a novel modified plane- and angle-sweeping strategy. In the proposed dense depth estimation technique, the whole scene is assumed to be region-wise planar and 3D models of these plane patches are estimated by a greedy-search algorithm that also considers visibility constraint. In order to refine the depth maps and relax the planarity assumption of the scene, at the final step, two refinement techniques that are based on region splitting and pixel-based optimization via Belief Propagation 62H[32] are also applied. Finally, the image segmentation algorithm is extended to object segmentation in multi-view video with the additional depth and optical flow information. Optical flow estimation is obtained via two different methods, KLT tracker and region-based block matching and the comparisons between these methods are performed. The experimental results indicate an improvement for the segmentation performance by the usage of depth and motion information.

Approximate Nearest-Neighbour Field (ANNF\ maps between two related images are commonly used by computer vision and graphics community for image editing, completion, retargetting and denoising. In this work we generalize ANNF computation to unrelated image pairs. For accurate ANNF map computation we propose Feature Match, in which the low-dimensional features approximate image patches along with global colour adaptation. Unlike existing approaches, the proposed algorithm does not assume any relation between image pairs and thus generalises ANNF maps to any unrelated image pairs. This generalization enables ANNF approach to handle a wider range of vision applications more efficiently. The following is a brief description of the applications developed using the proposed Feature Match framework. The first application addresses the problem of detecting the optic disk from retinal images. The combination of ANNF maps and salient properties of optic disks leads to an efficient optic disk detector that does not require tedious training or parameter tuning. The proposed approach is evaluated on many publicly available datasets and an average detection accuracy of 99% is achieved with computation time of 0.2s per image. The second application aims to super-resolve a given synthetic image using a single source image as dictionary, avoiding the expensive training involved in conventional approaches. In the third application, we make use of ANNF maps to accurately propagate labels across video for segmenting video objects. The proposed approach outperforms the state-of-the-art on the widely used benchmark SegTrack dataset. In the fourth application, ANNF maps obtained between two consecutive frames of video are enhanced for estimating sub-pixel accurate optical flow, a critical step in many vision applications. Finally a summary of the framework for various possible applications like image encryption, scene segmentation etc. is provided.