Segmentation and symmetry measure for image analysis

In this thesis, four topics related to image analysis are investigated: filtering, color segmentation, contour detection, and symmetry quantification for shape, color, and texture information. The three first topics are essential to detect efficiently the objects of interest present in the image, and to allow their analysis and classification. The use of features such as the degree of symmetry are necessary to this application. Although the combination of different features is usually necessary for an efficient classification, we want to focus on a unique feature and fully exploit it. In the framework of image filtering, we emphasize techniques which preserve the location and geometry of contours. Nonlinear isotropic filtering is therefore deeply investigated. This technique is inspired from the diffusion of heat in matter, with a nonlinear component that is based on the local gradient and that controls the diffusion. A unique threshold is used to limit and even stop the diffusion through the strongest transitions, which should then correspond to the object boundaries. This technique is particularly noise resistant and preserves the curvature of contours, which avoids rounding effects such as those encountered when using Gaussian low-pass filtering. Clustering techniques for color segmentation are investigated next. In particular, techniques using fuzzy sets theory are exploited, leading to the development of a new clustering scheme called orientation sensitive fuzzy c-means, whose main characteristic is to take into account the cluster orientation. This technique is therefore especially adapted to Gaussian and elliptic shaped mixtures. The clustering is performed in a two-dimensional histogram, which is computed with the two principal components of the Karhunen-Loève transform of the color frames. The number of clusters and the cluster center location must be obtained as well. The selection is based on geometrical features computed for every maximum in the two-dimensional histogram. In this thesis, the contour detection aims at the extraction of topology independent contours. Active contours, also called snakes, and their geometrical counterpart, geodesic active contours, are investigated for this purpose. These different techniques are presented and illustrated. Finally a new approach using a multi-components nonlinear isotropic diffusion applied to color images and morphological flooding is presented. This technique allows the extraction of a set of contour lines which are considered as contour candidates. The selection of valid contours is then done by global or local minimization of an energy functional. The last topic investigated in this thesis is the quantification of symmetry. Special care is given to its application to almost symmetrical and asymmetrical objects by having an optimization approach. Different techniques are developed to this end, using multi-resolution approaches, genetic algorithms, and self-organizing maps. The data is assumed to be vector-valued, such as are color and texture descriptors. Finally, the ability to detect symmetry axes in noisy environments is proven and illustrated. These different techniques are applied to the analysis or dermatoscopic images for diagnosis purpose. This type of images is used by dermatologists to diagnose skin cancer. The efficiency of the segmentation and contour detection techniques combined with the nonlinear isotropic filtering is illustrated. The improved separation between benign and malignant lesions obtained with our symmetry quantification approach, when compared to more classical techniques, is demonstrated using a simple linear classifier. We want to show in this study that the different diagnostic features, like the symmetry, must be fully investigated to design an efficient computer aided diagnosis system.