Segmentation of Dark Foreground Objects by Maximum Non–Extensive Entropy Partitioning

In this paper, we investigate the segmentation of dark foreground objects in a relatively bright background clutter by maximum non-extensive entropy partitioning of the image histogram. The non-linear, non-extensive entropy with Gaussian gain, which is a regularity indicator, is embedded in the maximum entropy thresholding framework, in the background of the recent work in Susan et al. (2016. Sadhana 41: 1393) on segregating facial intensities related to emotions by iterative maximum entropy partitioning. The non-linearity of this entropy ensures that out of the two partitions of the image histogram, the darkest shades in the image approximate very finely a uniform probability distribution and are separated out effectively from the brighter shades in the image that coarsely approximate a uniform distribution. This definition of the point of maximum entropy brought about by the non-linearity of the Gaussian curve targets and segments out the darkest shades pertaining to the foreground object in a more efficient manner than other thresholding techniques. Comparisons with the existing entropic thresholding schemes on test instances from a benchmark object segmentation dataset confirm the utility and efficiency of our method.