1. INTRODUCTION
Probabilistic graphical model, an interplay between probability and graph theory, provides a natural solution for uncertain and complex engineering problems [1]. Markov random field (MRF) [2] is a popular undirected graphical model used in the physics and vision field. This paper provides a new unsupervised MRF method based on superpixels to solve the region segmentation problem. Image segmentation is a fundamental problem in both image processing and computer vision. MRF based segmentation methods are very appealing for various reasons. It incorporates both prior knowledge and local spatial relationship. It has performance evaluation in a natural way. MRF methods based on pixels or regular shape neighbors are widely explored in theoretical and practical research [3]. Pixel based method is computationally complex and intractable. Recently, several superpixel based methods for supervised learning [4] [5] [6] are proposed. One supervised superpixel based MRF for 3D reconstruction [4] uses 4 neighbor structure feature extraction. In [5], superpixel surface layout learning does not employ MRF and uses location as a feature. The irregular labeling scheme in [6] only considers the two neighbor pairwise structure, uses only gray level as feature and has no numerical comparison. Although superpixel concept is widely used in computer vision, it has not been applied to unsupervised learning.