François Mendels
This thesis is dedicated to the study of a new framework linking image segmentation and shape recognition processes in the image processing domain. Image segmentation process aims at separating an image into homogenous regions, the homogeneity measure depending on the feature chosen, such as color, texture. The goal of a shape recognition process is to identifying object(s) present in an image whose shape is similar to a chosen model, with respect to e.g. given rigid transformations that do not modify the intrinsic nature of the shape. As these kinds of process are nowadays integrated in industrial applications for automation purposes (e.g. detecting a defect on manufactured tools), it is necessary to have flexible, and evolving processes that remain robust in adverse conditions. These two processes are usually realized separately, thus one does not benefit from both approach advantages, and thus, can not automatically correct each other errors. In this work, we wish to add a priori information, such as a given shape of an object, into a segmentation framework, here the active contour framework. Using this framework, we focus on segmenting multiple objects belonging to the same shape space. Up to now, the models including a shape prior in the active contour models, by modifying their energy functional, are unable to solve this problem directly. We choose here to integrate the shape prior in two different ways: we modify the external field, in which the active contour propagates, or we use the a priori information as new initial conditions. In order to obtain a shape prior having both local and global conditions, we introduce a new shape descriptor relying on an analytical low-level representation of images, based on a generic framework using sparse representation algorithms. This continuous shape description is obtained by a decomposition process, here the Matching Pursuit (MP) algorithm using an affine invariant dictionary of basis functions. This description is then used in the shape recognition process. In the learning phase, a template object is decomposed, and the extracted subset of basis functions, called meta-atom, gives the description of our object. We then extend naturally this description into the linear scale-space using the definition of our basis functions, and thus bringing a more general representation of our object. We use this enhanced description as a predefined dictionary of the object to conduct an MP-based shape recognition (MPSR) task into the linear scale-space: we extract iteratively each object belonging to the given shape space. We introduce two different methods to realize this MPSR process in the linear scale-space: a coarse-to-fine approach where the solutions found at higher scales are propagated to lower scales, and a stack scale-space approach where the extraction is realized at all scales at once. The introduction of the scale-space approaches improves the robustness of our method, and permits to avoid local minima problems encountered when minimizing a non-convex energy function. The shape prior is build as the sum of the solutions resulting from the MPSR process. We propose three different ways to introduce the a priori information in the active contour, reflecting three different segmentation goals. At first, we consider the shape prior as the new external field, thus this enables us to deal with occlusion problems. The second proposal combines the influence from the prior information and the original external field, and lead to let the active contour converge towards real boundaries close to the shape prior. The third proposal sums the variable actions of the shape prior and the external field: in this case, the shape prior is strong at the beginning of the process, and it is relaxed during the process to let the active contour converge to the existing boundaries present in the target image. This enables us to extract objects having large deformations when compared to the shape model. We then show results for the detection and segmentation of complex synthetic shapes, as well as natural (aerial and medical) images, for all these approaches, in coherence with the expected segmentation goals. The global approaches remain resistant to noise presence.
EPFL2004
