Fast Marching Techniques for Visual Grouping & Segmentation

Abstract

A new region growing method is proposed for segmenting images. The region boundaries are formulated as level sets and the pixel labeling process is implemented using a new multi-label fast marching algorithm. The region contours are propagated with a velocity proportional to the a posteriori probability of the respective label. Statistical tests are performed to generate the initially labeled sets. Any image feature, given it is semantically relevant, can be considered for the segmentation process. Illustrations are given for combined luminance, chrominance and texture classification and segmentation in natural scenes. Moving object extraction based on change detection is also considered, which is performed as a two-label classification.