Contour and Region-Based Image Segmentation

Abstract

One of the most complex tasks in computer vision is segmentation. Segmentation can be roughly defined as optimally segregating the foreground from the background, or by finding the optimal partition of the image into its constituent parts. Here optimal segregation means that pixels (or blocks in the case of textures) in the foreground region share common statistics. These statistics should be significantly different from those corresponding to the background. In this context, active polygons models provide a discriminative mechanism for the segregation task. We will show that Jensen–Shannon (JS) divergence can efficiently drive such mechanism. Also, the maximum entropy (ME) principle is involved in the estimation of the intensity distribution of the foreground.

It is desirable that the segmentation process achieves good results (compared to the ones obtained by humans) without any supervision. However, such unsupervision only works in limited settings. For instance, in medical image segmentation, it is possible to find the contour that separates a given organ in which the physician is interested. This can be done with a low degree of supervision if one exploits the IT principle of minimum description length (MDL). It is then possible to find the best contour, both in terms of organ fitting and minimal contour complexity. IT inspires methods for finding the best contour both in terms of segregation and minimal complexity (the minimum description length principle).