The Contracting Curve Density Algorithm and its Application to Model-based Image Segmentation

In this paper we address the problem of model-based image segmentation by fitting deformable models to the image data. From uncertain a priori knowledge of the model parameters an initial probability distribution of the model edge in the image is obtained. From the vicinity of the surmised edge local statistics are learned for both sides of the edge. These local statistics provide locally adapted criteria to distinguish the two sides of the edge even in the presence of spatially changing properties such as texture, shading, or color. Based on the local statistics the model parameters are iteratively refined using a MAP estimation. Experiments with RGB images show that the method is capable of achieving high subpixel accuracy and robustness even in the presence of texture, shading, clutter, and partial occlusion.