Quantization from Bayes factors with application to multilevel thresholding

Abstract

We are concerned with the optimal selection of multiple thresholds in image analysis. We propose the use of the Bayes information criterion, a minimal information measure, for this and illustrate its use in practical cases. Applications of multiple threshold selection of interest to us include the closely related problems of (i) quantization for lossy encoding, and (ii) segmentation. Our examples relate to segmentation as a post-processing phase in edge detection.