Dans cet article, nous proposons une méthode originale pour incorporer un a priori de forme dans un modèle de contours actifs basé région afin d'améliorer sa robustesse aux similitudes, bruit et occultations. Nous définissons un a priori de forme à partir du recalage des fonctions level set associées au contour actif et une forme de référence. Le recalage que nous proposons se base sur la corrélation de phase par la transformée de Fourier-Mellin analytique (TFMA). Cette représentation, dédiée aux images à niveaux de gris, permet de gérer simultanément plusieurs objets. Nous illustrons expérimentalement les capacités de ce nouvel a priori de forme à contraindre l'évolution du contour actif vers une forme cible. Enfin, nous mettons en évidence, sur des images de synthèse et réelles, son apport pour la segmentation d'images en présence de similitudes, d'occultations et de bruits. ABSTRACT.
In this paper, we propose new method to incorporate geometric shape prior into region-based active contours in order to improve its robustness to noise and occlusions. The proposed shape prior is defined after the registration of the level set functions associated with the active contour and a reference shape. The used registration method is based on phase correlation by the Analytical Fourier-Mellin Transform (AFMT). This representation, dedicated to gray levels images, makes it possible to manage several objects simultaneously. Experimental results show the ability of the proposed geometric shape prior to constrain an evolving curve towards a target shape. We highlight on synthetic and real images, the benefit of the new shape prior on segmentation results, in presence of occlusions and noise. MOTS-CLÉS : contours actifs, a priori de forme, transformée de Fourier-Mellin analytique. Extended abstract Active contours have been introduced in 1988 (Kass et al., 1988). The principle is to move a curve iteratively minimizing energy functional. The minimum is reached at object boundaries. These methods can be classified into two families: parametric and geometric active contours. The first family, called also snakes, uses an explicit representation of the contours while the second one uses an implicit representation of the front via Level Set approach. These models typically manage the evolution of the active contour based on local information of the image (gray level). The lack of global information about the target object prevents these approaches to be robust in presence of textured background, occlusions or even noise. Several studies have proposed to introduce prior knowledge on the shape to detect into the active contour model. In the context of statistical shape prior, Leventon et al. (2000) proposed to associate a statistical model on the learned shape to the geodesic active contours (Caselles et al., 1997). Chen et al. (2001) defined an energy functional based on the quadratic distance between the evolving curve and the average shapes of the target object after alignment. This term is then incorporated into the geodesic active contours. Foulonneau et al. (2004) introduced an additional geometric shape prior into region-based active contours. Prior knowledge is defined as a distance between shapes descriptors based on Legendre moments of the characteristic function. An extension of this work in case of affine transformation is performed in (Foulonneau et al., 2006) and in (Foulonneau et al., 2009), a multi-references shape prior is presented. Charmi et al. (2008) introduced geometric shape prior into the snake model. A set of complete and locally stable invariants to Euclidean transformations (Ghorbel. 1998) is used to define new force which makes the snake overcome some well-known problems. In (Charmi et al., 2010), we defined new geometric shape prior for region-based active contours (Chan, Vese, 2001). The new added term was based on the property of signed distance function associated with the evolving contour which assigns negative values for points inside the contour and positive values for those outside. In fact, given two level set functions associated with the active contour and the template, variability between shapes can be formulated using the Heaviside of the product function of these two level set functions. Hence the proposed shape prior is the integral of this function on the image domain. This term is then incorporated into the evolution's equation of the active contour. If one takes a reference shape which is not necessarily defined in the image reference, it is necessary to apply a transformation to align it with the shape to segment (rotation, translation, scaling factor). Then, we used a shape alignment method based on Fourier descriptors. This shape prior introduced into region-based active contours has been successful in case of single object in the image in presence of noise and occlusion. It is well known that the level set approach solves the problem of topology changing of the snake model. However, shape prior in several works (Leventon et al. (2000), Chen et al. (2001), Chan et Zhu (2005), Fang et Chan (2007) and Charmi Contours actifs avec a priori de forme basé sur la TFMA 125 et al. (2010)) based on contours alignment constrain these approaches to segment only one object in the image. Thus our goal in this paper is to extend the work done in Charmi et al. (2010) to manage the case of several objects that may be partially occluded and possibly noisy. We were based on the property of distance maps of level set functions and associated binary image for every distance map. Then the problem amounts to the registration of these images. We used the method of phase correlation in Fourier space that is appropriate to estimate the translation vector and phase correlation in the space of Fourier-Mellin for estimating the rotation and the scaling factor knowing that the Fourier-Mellin transform applied to grayscale images is a mathematical tool known for its performance in objects description and features recognition. It was also pointed out that numerical estimation of the Mellin integral brings up crucial difficulties. A solution for the convergence of the integral was given in (Ghorbel, 1994) by using the analytical Fourier-Mellin transform (AFMT). In (Derrode, Ghorbel, 2001), three approximations of the AFMT were proposed. We adopt the fast algorithm based on fast Fourier transform (FFT) and log-polar sampling of the image. Hence to introduce the shape prior we proceed as follows:-we start by segmenting the target object with the Chan and Vese's model without prior knowledge;-then, after convergence of the evolving contour, we register the binary images associated with the level set functions of the evolving contour and the template;-having the parameter of the rigid transformation between shapes, we calculate the proposed shape prior;-finally, the proposed model evolves under Chan-Vese and the prior knowledge terms with big weight assigned to the shape prior energy to constrain the active contour to be similar to the template; Experiments have shown the ability of the new added term to improve the robustness of the segmentation process in presence of textured background, missing parts and partial occlusions of the target object. The addition of shape prior has not increased significantly the execution time of the algorithm given that the proposed approach does the registration only once and it is done by the Fast Fourier Transform (FFT2) unlike Foulonneau et al. (2004) and Charmi et al. (2008) where at each iteration, shape descriptors are calculated for a given order. As future perspectives, we are working on extending this approach to more general transformations such as affine transformations and manage the case where many references are available and thus the model must be able to choose the most suitable shape according to the evolving contours.