Mesures statistiques non-paramétriques pour la segmentation d'images et de vidéos et minimisation par contours actifs

摘要

Image and video segmentation consists in the partitioning of an image into objects of interest and background. When using active contours in an variational framework, the difficulty is to define an appropriate segmentation criterion. This criterion is then differentiated using shape gradients, in order to obtain the evolution equation of the active contour. Often this criterion depends on image features and makes an assumption on the distribution of such features. For example, considering a function of the intensity mean as a criterion is equivalent to making a Gaussian assumption on the distribution of the intensity. In this work, we propose to get rid of such assumptions by approximating actual distributions. We use a non-parametric kernel-based estimator. We propose different criteria coming from information theory, such as entropy, to segment zones with limited intensity variations. In order to take into account several channels like color channels, two alternatives are proposed : joint entropy and mutual information. When some a priori is available, the Kullback-Leibler divergence is used to minimize a distance between a reference distribution and the distribution of the current region. To segment moving objects in video sequences, the joint entropy is used. A first approach consists in computing the optical flow and minimizing the joint entropy of its components. A second approach consists in jointly estimating the motion and segmenting moving objects by minimizing the joint entropy of a residual and the image intensity.