Unsupervised segmentation based on connectivity analysis

In this paper, we present an original unsupervised segmentation scheme which is based on connectivity analysis. This scheme splits a grey level image into different sets of connected pixels with homogeneous grey levels. Our approach is based on a multiscale analysis of a triangular table called “connectivity degrees pyramid”. First, we extract the connectivity degrees´ local maxima on each line of the pyramid and we consider that the local maxima of the base line correspond to candidate classes. For each candidate class, we construct its fingerprint thanks to the tracking of its local maxima in the higher lines. We then construct the classes by selecting the candidate classes and by a multiscale analysis of their fingerprint. Finally, we assign the pixels to their respective class according to the grey level cooccurrences. The efficiency of our approach is illustrated by the segmentation of a well known image