Region Extraction Using Competition of Multiple Active Contour Models

In most region extraction methods using the Active Contour Model (ACM, Snakes), the entire area of a target object is treated as a single region, and an ACM is deformed locally by image properties around the contour. For these reasons, it is difficult for these methods to extract accurately an object area that has intricate distributions of image properties. In response to this difficulty, to introduce wide-ranging image properties into the extraction process, it has been proposed to control an ACM by the statistical characteristics of image properties in the object and background regions. However, this approach is unsuitable when image properties vary widely with location in each region. In this paper, to overcome this problem, a novel region extraction method is proposed. The proposed method makes multiple ACMs compete with each other and each ACM extract a subregion of uniform image properties, and as a result, the method extracts the entire area of an object as a set of several subregions. In the proposed method, first, a few initial curves are set in the object and background regions. Each initial curve is divided into several segments, and multiple initial ACMs are made by dilating these segments. Second, at each ACM, the density function of image properties is estimated, and for each control point on a contour, the likelihood with respect to every ACM region is determined. Based on this likelihood, each ACM is controlled and multiple ACMs compete with each other.