Semantic image segmentation with Markov Random Fields

Abstract

In the recent studies image segmentation and object recognition are handled cooperatively. Majority of those studies employ supervised or semi-supervised training by providing labels. However, providing labeling is too laborious. For this reason, we propose using prior knowledge on domain information instead of class labels. Given the domain knowledge the system detects domain invariants in the image. By means of detecting domain invariants, it obtains an initial segmentation of the image. This initial segmentation is further improved by a Markov Random Field based segmentation method. So, the proposed method consists of two parts; in the first part, an initial segmentation is obtained by detecting the domain invariant(s) in the image, in the second part, the initial segmentation is improved by means of a Markov Random Field based segmentation algorithm.