Binary Segmentation-Based Image Coding

The Minimal Message Length (MML) Principle has been proposed by Wallace as an objective method for comparing diierent models which are seeking to explain a body of data. We apply MML to the problem of comparing diierent bi-level segmentations of an image. This is done by generalising the concept of an MML two-part message to the problem domain of images. The model part of the two-part message is a binary image which identiies the bi-level segmen-tation while the second part of the message contains the information which combined with the binary image allows exact reconstruction of the original image. Experimental results are presented for diierent segmentation schemes such as binary thresholding, k-means clustering, an MML classiication scheme-snob and fuzzy c-means. The application of ltering techniques before thresh-olding is also considered.