Study On Image Segmentation Based On Gaussian Mixture Model

Image segmentation is an important part of computer vision and widely used in image recognition system. After decades of research, many scholars have proposed a variety of methods and theories about image segmentation. Image segmentation technique based on models is a hot spot in these methods and techniques, Gaussian Mixture Model (GMM) is an effective and well-known model in these model-based techniques. Cluster analysis methods based on Gaussian Mixture Model can resolve complex content and uncertainty classification questions about image segmentation. But prior probability distribution based on Gaussian Mixture Model without introducing spatial relationships between neighboring pixels, resulting in segmentation results which got from Gaussian Mixture Model for image segmentation are more sensitive to noise. Therefore, Spatially Variant Finite Mixture Model (SVFMM) firstly proposed the introduction of spatial relationships between neighboring pixels to prior probability distribution of mixture model, so that mixture model has better noise immunity. To improve mixture model’s noise immunity many scholars improved the SVFMM. In order to overcome the problem of noise sensitive to GMM, there are two ways to make spatial relationship constraints prior probability distribution on the GMMFirstly, study on image segmentation method based on Gaussian Mixture Model in this paper, as well as introduce the Markov Random Field (MRF) and introduction of Markov Random Field to prior probability distribution of Gaussian Mixture Model. And then introduce two methods for parameter estimation, Expectation-Maximization (EM) algorithm and Gradient Descent Algorithm. Also describe advantages and disadvantages of Gaussian Mixture Model for image segmentation, And these can be used as the basis of the further study.Secondly, we study a method which is an extension of the Gaussian Mixture Model, it is a method of the first kind of the prior probability constraint in the Gaussian Mixture Model. We proposed a method of constructs prior probability distribution replacing each pixel value in an image with the average value of its neighbors(including itself), based on the thought that the distribution of prior probability of adjacent pixels are more likely to be tended to be the same or similar. The mean of the neighborhood of the weight function is applied to the prior probability distribution, improve mixture model’s noise immunity, and use the Gradient Descent Algorithm to estimate parameters, and gives the pseudo code of the algorithm.Finally, we propose a Gaussian Mixture Model based on Gaussian kernel function’s prior probability, the model get the Gaussian kernel function and weight function that each pixel belonging to each category of the extension of the Gaussian Mixture Model together constraint the priori probability. The model and GMM were used to experiment with the synthetic images, and the results showed that the model had better noise immunity.