HAL is a multi-disciplinary open access archive for the deposit and dissemination of scientific research documents, whether they are published or not. The documents may come from teaching and research institutions in France or abroad, or from public or private research centers. L’archive ouverte pluridisciplinaire HAL, est destinée au dépôt et à la diffusion de documents scientifiques de niveau recherche, publiés ou non, émanant des établissements d’enseignement et de recherche français ou étrangers, des laboratoires publics ou privés. Extraction de motifs spatio-temporels : co-localisations, séquences et graphes dynamiques attribués Frédéric Flouvat

Extraction de motifs spatio-temporels : co-localisations, séquences et graphes dynamiques attribués

RX Detector is recognized as the benchmark algorithm for image anomaly detection, however it presents known limitations, namely the dependence over the image following a multivariate Gaussian model, the estimation and inversion of a highly dimensional covariance matrix and the inability to effectively include spatial awareness in its evaluation. In this work a novel graph-based solution to the image anomaly detection problem is proposed; leveraging on the Graph Fourier Transform, we are able to overcome some of RX Detector's limitations while reducing computational cost at the same time. Tests over both hyperspectral and medical images, using both synthetic and real anomalies, prove the proposed technique is able to obtain significant gains over RX Detector performance.

Graph-based Image Anomaly Detection.

We propose a method for semantic image segmentation, combining a deep neural network and spatial relationships between image regions, encoded in a graph representation of the scene. Our proposal is based on inexact graph matching, applied to the output of a deep neural network. The proposed method is evaluated on a public dataset used for segmentation of images of faces. Preliminary results show that, in terms of IoU of region bounding boxes, the use of spatial relationships lead to an improvement of 2.4% in average, and up to 24.4% for some regions.

https://hal.archives-ouvertes.fr/hal-02882043/document

Méthode d’analyse sémantique d’images combinant apprentissage profond et relations structurelles par appariement de graphes

Image recovery and segmentation are always the fundamental tasks in image processing field, because of their so many contributions in practical applications. As in the past ten years, variational methods have achieved a great success on these two issues, in this thesis, we continue to work on proposing several new variational approaches for restoring and segmenting an image. This thesis contains two parts. The first part addresses recovering an image and the second part emphasizes on segmenting. Along with the wide utilization of magnetic resonance imaging (MRI) technique, we particularly deal with blurry images corrupted by Rician noise. In chapter 1, two new convex variational models for recovering an image corrupted by Rician noise with blur are presented. These two models are motivated by the non-convex maximum-a-posteriori (MAP) model proposed in the prior papers. In the first method, we use an approximation item to the zero order of the modified Bessel function in the MAP model and add an entropylike item to obtain a convex model. Through studying on the statistical properties of Rician noise, we bring up a strictly convex model by adding an additional data-fidelity term in the MAP model in the second method. Primal-dual methods are applied to solve the models. The simulation outcomes show that our models outperform some existed effective models in both recovery image quality and computational time. Cone beam CT (CBCT) is routinely applied in image guided radiation therapy (IGRT) to help patient setup. Its imaging dose, however, is still a concern, limiting its wide applications. It has been an active research topic to develop novel technologies for radiation dose reduction. In chapter 2, we propose an improvement of practical CBCT dose control scheme temporal non-local means (TNLM) scheme for IGRT. We denoise the scanned image with low dose by using the previous images as prior knowledge. We combine deformation image registration and TNLM. Different from the TNLM, in the new method, for each pixel, the search range is not fixed, but based on the motion vector between the prior image and the obtained image. By doing this, it is easy to find the similar pixels in the previous images, but also can reduce the computational time since it does not need large search windows. The phantom and patient studies illuminate that the new method outperforms the original one in both image quality and computational time. In the second part, we present a two-stage method for segmenting an image corrupted by blur and Rician noise. The method is motivated by the two-stage segmentation method developed by the authors in 2013 and restoration method for images with Rician noise. First, based on the statistical properties of Rician noise, we present a new convex variant of the modified Mumford-Shah model to get the smooth cartoon part u of the image. Then, we cluster the cartoon u into different parts to obtain the final contour of different phases of the image. Moreover, u from the first stage is unique because of the convexity of the new model, and it needs to be computed only once whenever the thresholds and the number of the phases K in the second stage change. We implement the simulation on the synthetic and real images to show that our model outperforms some existed segmentation models in both precision and computational time.

Variational approaches in image recovery and segmentation

The latest generation of aerial- and satellite-based imaging sensors acquires huge volumes
of Earth’s images with high spatial, spectral and temporal resolution, which open the door
to a large range of important applications, such as the monitoring of natural disasters, the
planning of urban environments and precision agriculture. In order to fully exploit the
potential offered by these sensors, there is a need to develop accurate and time-efficient
mathematical models and algorithms for spectral-spatial analysis of the recorded high-
resolution data.
The main goal of my research is to develop learning approaches, which would help
to automatically interpret, or classify, remote sensing images. This manuscript presents
several strategies I have explored for this purpose, varying from the use of strong shape
priors to detect objects, regularization of classification probabilities on the image graphs,
and up to the use of convolutional neural network models capable to learn deep hierar-
chical contextual features.
The experimental results on diverse benchmarks of images and image time series show
the competitiveness of the developed methods when compared to the state-of-the-art ap-
proaches. In particular, we have recently created large-scale classification benchmark
of aerial images and have demonstrated that the modern deep learning-based methods
succeed in generalizing to the dissimilar urban settlements around the Earth. This opens
new exciting perspectives towards designing systems which would be able to automati-
cally update world-scale maps from remote sensing data.

Learning Approaches for Remote Sensing Image Classification

Graph Theory Concepts and Definitions Used in Image Processing and Analysis, O. Lezoray and L. Grady Introduction Basic Graph Theory Graph Representation Paths, Trees, and Connectivity Graph Models in Image Processing and Analysis Graph Cuts-Combinatorial Optimization in Vision, H. Ishikawa Introduction Markov Random Field Basic Graph Cuts: Binary Labels Multi-Label Minimization Examples Higher-Order Models in Computer Vision, P. Kohli and C. Rother Introduction Higher-Order Random Fields Patch and Region-Based Potentials Relating Appearance Models and Region-Based Potentials Global Potentials Maximum a Posteriori Inference A Parametric Maximum Flow Approach for Discrete Total Variation Regularization, A. Chambolle and J. Darbon Introduction Idea of the approach Numerical Computations Applications Targeted Image Segmentation Using Graph Methods, L. Grady The Regularization of Targeted Image Segmentation Target Specification Conclusion A Short Tour of Mathematical Morphology on Edge and Vertex Weighted Graphs, L. Najman and F. Meyer Introduction Graphs and lattices Neighborhood Operations on Graphs Filters Connected Operators and Filtering with the Component Tree Watershed Cuts MSF Cut Hierarchy and Saliency Maps Optimization and the Power Watershed Partial Difference Equations on Graphs for Local and Nonlocal Image Processing, A. Elmoataz, O. Lezoray, V.-T. Ta, and S. Bougleux Introduction Difference Operators on Weighted Graphs Construction of Weighted Graphs p-Laplacian Regularization on Graphs Examples Image Denoising with Nonlocal Spectral Graph Wavelets, D.K. Hammond, L. Jacques, and P. Vandergheynst Introduction Spectral Graph Wavelet Transform Nonlocal Image Graph Hybrid Local/Nonlocal Image Graph Scaled Laplacian Model Applications to Image Denoising Conclusions Acknowledgments Image and Video Matting, J. Wang Introduction Graph Construction for Image Matting Solving Image Matting Graphs Data Set Video Matting Optimal Simultaneous Multisurface and Multiobject Image Segmentation, X. Wu, M.K. Garvin, and M. Sonka Introduction Motivation and Problem Description Methods for Graph-Based Image Segmentation Case Studies Conclusion Acknowledgments Hierarchical Graph Encodings, L. Brun and W. Kropatsch Introduction Regular Pyramids Irregular Pyramids Parallel construction schemes Irregular Pyramids and Image properties Graph-Based Dimensionality Reduction, J.A. Lee and M. Verleysen Summary Introduction Classical methods Nonlinearity through Graphs Graph-Based Distances Graph-Based Similarities Graph embedding Examples and comparisons Graph Edit Distance-Theory, Algorithms, and Applications, M. Ferrer and H. Bunke Introduction Definitions and Graph Matching Theoretical Aspects of GED GED Computation Applications of GED The Role of Graphs in Matching Shapes and in Categorization, B. Kimia Introduction Using Shock Graphs for Shape Matching Using Proximity Graphs for Categorization Conclusion Acknowledgment 3D Shape Registration Using Spectral Graph Embedding and Probabilistic Matching, A. Sharma, R. Horaud, and D. Mateus Introduction Graph Matrices Spectral Graph Isomorphism Graph Embedding and Dimensionality Reduction Spectral Shape Matching Experiments and Results Discussion Appendix: Permutation and Doubly- stochastic Matrices Appendix: The Frobenius Norm Appendix: Spectral Properties of the Normalized Laplacian Modeling Images with Undirected Graphical Models, M.F. Tappen Introduction Background Graphical Models for Modeling Image Patches Pixel-Based Graphical Models Inference in Graphical Models Learning in Undirected Graphical Models Tree-Walk Kernels for Computer Vision, Z. Harchaoui and F. Bach Introduction Tree-Walk Kernels as Graph Kernels The Region Adjacency Graph Kernel as a Tree-Walk Kernel The Point Cloud Kernel as a Tree-Walk Kernel Experimental Results Conclusion Acknowledgments

Qiang Sun

Papers

图像处理和分析的图模型：理论与应用 (Image Processing and Analysis with Graphs: Theory and Practice)