Image Segmentation Through Contextual Clustering

Several interesting strategies are adopted by the well-known Pappas clustering algorithm to segment smooth images. These include exploitation of contextual information to model both class conditional densities and a priori knowledge in a Bayesian framework. Deficiencies of this algorithm are that: i) it removes from the scene any genuine but small region; and ii) its feature-preserving capability largely depends on a user-defined smoothing parameter. This parameter is equivalent to a clique potential of a Markov Random Field model employed to capture known stochastic components of the labeled scene. In this paper a modified version of the Pappas segmentation algorithm is proposed to process smooth and noiseless images requiring enhanced pattern-preserving capability. In the proposed algorithm: iii) no spatial continuity in pixel labeling is enforced to capture known stochastic components of the labeled scene; iv) local intensity parameters, pixel labels, and global intensity parameters are estimated in sequence; and v) if no local intensity average is available to model one category in the neighborhood of a given pixel, then global statistics are employed to determine whether that category is the one closest to pixel data. Results show that our contextual algorithm can be employed: vi) in cascade to any noncontextual (pixel-wise) hard c-means clustering algorithm to enhance detection of small image features; and vii) as the initialization stage of any crisp and iterative segmentation algorithm requiring priors to be neglected on earlier iterations (such as the Iterative Conditional Modes algorithm).

ICSI Technical Report TR-98-009

Technical Report