An MML-based criterion for comparing image segmentations

Usually the performance of segmentation methods is judged by comparison to manual methods and subjective criteria. There are no generally accepted objective criteria for choosing a suitable segmentation currently. Wallace introduced the Minimum Message Length (MML) Principle as an objective method for comparing different models for explaining a body of data. In this paper, we use the ML principle to provide an objective criterion for the comparison of different image segmentations of the same image. This is done by constructing an MML two-part message for the image. Experimen tal results are presented for different segmentation schemes based on thresholding such as histogram thresholding, k-means clustering, an MML classifcation scheme snob, and fuzzy c-meaas. The application of a filtering technique before thresholding is also considered.