Extended Markov Random Fields for Predictive Image Segmentation

Abstract

Since the 1970s, there has been increasing interest in the use of Markov Random Fields (MRFs) as models to aid in the segmentation of noisy or degraded digital images. MRFs can make up for deficiencies in observed information by adding a-priori knowledge to the image interpretation process in the form of models of spatial interaction between neighbouring pixels. In data fusion problems, interaction might also be assumed between corresponding pixels in two different kinds of image of the same scene. Alternatively, temporal interaction might be assumed between corresponding pixels in consecutive frames of a video sequence. In object tracking or robotic navigation problems, a similar relationship may exist between pixels of an observed image and those of a predicted image, derived from models of the motion and scene. In all of these cases the MRF model can be extended to incorporate this additional knowledge. This paper explains the theory of Extended-Markov Random Field (E-MRF) segmentation techniques, surveys the research which has been crucial to their development and presents results from new work in this area with an application to robotic vision in conditions of extremely poor visibility.