Using spatial information as an aid to maximum entropy image threshold selection

Abstract Many image grey-level thresholding methods based on the theory of maximum entropy have been proposed in the past. However, the exact definition of what is meant by “image entropy” has varied considerably, with measures based on the image itself, its histogram or other related distributions being proposed. Most of these ignore the spatial component of the image, assuming pixels within the image to be independent of each other. This is convenient, but it is also counter-intuitive. The inclusion of context-related information can be a simple matter, as is demonstrated in this article. A common measure of image entropy is based on the notion of viewing the image itself as a probability distribution (often referred to as the “monkey model” of the image). Direct application of this measure as a criterion function for grey-level threshold selection yields generally unsatisfactory results. With the inclusion of spatial information in the entropy measure, the results improve dramatically.