Contextual clustering for image segmentation

The unsupervised Pappas adaptive clustering (PAC) algorithm is a well-known Bayesian and contextual procedure for pixel labeling. It applies only to piecewise constant or slowly varying intensity images that may be corrupted by an additive white Gaussian noise field independent of the scene. Interesting features of PAC include multiresolution implementation and adaptive estimation of spectral parameters in an iterative framework. Unfortunately, PAC removes from the scene any genuine but small region whatever the user-defined smoothing parameter may be. As a consequence, PAC's application domain is limited to providing sketches or caricatures of the original image. We present a modified PAC (MPAC) scheme centered on a novel class-conditional model, which employs local and global spectral estimates simultaneously. Results show that MPAC is superior to contextual PAC and stochastic expectation-maximization as well as to noncontextual (pixel-wise) clustering algorithms in detecting image details.