The Study And Application Of Image Segmentation Based On Finite Mixture Model

Image segmentation denotes a process by which a raw image is partitioned into nonoverlapping regions, and at the same time it is an important direction of digital image processing and the basic of machine vision. Image segmentation is the foundation of feature extraction and image understanding. Correct segmentation provides powerful evidence for information determination. As a carrier of information, images are inherently complex and of rich content. Although the traditional threshold segmentation, the segmentation method base on region growing, split and merge algorithm method show good performance simple Image, there remain some limitations on the applications. The essence of image segmentation is the clustering of image pixels. Mixture models have been successfully applied to image segmentation since it is easy to be implemented. However, the traditional finite mixture model cannot get good image segmentation results on noisy images due to the relationship between pixels is not considered. Markov random field model considers the spatial relationships between pixels have been successfully applied to image segmentation, which still remains some problems in the image segmentation. On the basis of the study of finite mixture model, this paper proposes the following innovations:1. Since traditional image segmentation methods are not effective in the case of image smearing with heavy-tailed noise, this paper presents an image segmentation method based on student-t distribution which has good capability of denoising. Compared with the Gaussian mixture model(GMM), the algorithm takes the spatial relationships between image pixels into account. Meanwhile, compared with the model based on the Markov random fields, the algorithm requires less parameters. So it is easier to be achieved. Gradient method is used to estimate model parameters instead of utilizing an expectation-maximization algorithm. The experimental results show that mixture model based on student-t distribution can effectively dispose the image smearing with heavy-tailed noise.2. When using the existed improved Gaussian mixture model in image segmentation, how to speed up the segmentation process is a significant research topic. Based on the latest noise-benefit EM algorithm, this paper speeds up the convergence speed of the existed improved Gaussian mixture model by adding artificial noise, which achieves the goal of speeding up image segmentation. Additive noise speeds up the average convergence of the EM algorithm to a local maximum of the likelihood surface when adding noise to meet the noise-benefit EM theorem. Improved Gaussian mixture model is a special case of the expectation-maximization algorithm. Therefore, noise-benefit EM theorem can be applied to improved Gaussian mixture model. Experimental results indicate that the proposed algorithm speeds up the convergence process in image segmentation, and the time complexity is decreased significantly.