Image Segmentation Based On Mutual Information And Markov Random Field

Image segmentation is a very important research field in the scope of image processing. It has extensive application and involves almost all fields such as image understanding, pattern recognition and image encoding, etc. Furthermore, research of image segmentation is conducted on all types of images. In its application in medicine, it is achieved to segment the brain magnetic resonance (MR) image into different tissue classes, especially gray matter (GM), white matter (WM), cerebrospinal fluid (CSF).Mutual information (MI) is a basic concept from Shannon information theory and has been widely used in image registration before. Recently, some segmentation methods based on MI are reported.In this paper a new segmentation method called DDC, based on difference of mutual information (dMI) and pixon, is proposed. Experiments demonstrate that dMI shows one kind of intrinsic relationship between the segmented image and the original one and so it can be used to well determine the number of clusters. Furthermore, medical images with lesions can be automatically and successfully segmented by DDC method.An optimization objective function of image segmentation depend on MI is introduced too in this paper. Based on that, a new segmentation method named MMS (Mutual information Maximized Segmentation) is proposed. In MMS method, simulated annealing is used for finding the global minimum. Experiments show that the segmentation by MMS has the more entropy value, which is equal to MI between the original image and its segmentation, than that by FCM.Markov random field (MRF) well describes the relationship between the current pixels and their neighbors, and as a result it has been gaining wide concerns and applications. The MRF model is a nice tool to encompass prior knowledge in the segmentation process, while image segmentation is frequently formulated as a Bayesian and MAP (maximum a posteriori) criterion is often used.In this paper, we propose a new image segmentation method based on tree-structured Markov random field (TS-MRF) and fuzzy multi-level logistic (MLL) model, where fuzziness has been introduced into TS-MRF and the calculation of potential function becomes more meticulous and more precise. Compared with the classical TS-MRF method, the quantitative errors are smaller under the same conditions when the new algorithm is used, while the computational time increases little. More interesting, this method may hint one simple and efficient way to make MRF-based prior (not limited in MLL model) more precise, i. e. making the MRF-based clique changed to be fuzzy clique by using posterior probability.At last, we propose a novel method to judge whether a segmentation step ought to stop or not, which is named "catalyst" segmentation method (CS). Primary experiments shows that CS is simple and useful.In the future work, we want to explore the application in segmentation of special goal, such as tumor lesion and the number of test images will be increased. Secondly, we think that fuzzy MLL can be looked as a generic model and try to use it in other scope beyond image segmentation. Thirdly, we will consummate the CS method.