Gaussian mixture model (GMM) is easily affected by noise, and Markov random field (MRF) model can well describe the spatial characteristics. The combination of the two is suitable for image segmentation with noise, but MRF model is prone to over segmentation. To solve this problem, an improved image segmentation algorithm based on adaptive weight coefficient was proposed, which can segment cerebrospinal fluid, gray matter and white matter from magnetic resonance imaging (MRI). Firstly, the K?means algorithm was used to obtain the initial segmentation results, and the Expectation?Maximization (EM) algorithm was used to estimate the parameters of GMM, and then the joint probability energy function of the pixel gray level of the image was obtained. Then, the adaptive weight coefficient was obtained by using the gray value, posterior probability and Euclidean distance of the center pixel and the neighboring pixels of the MRF neighborhood system, and the prior probability energy function was obtained by MRF. Finally, the final image segmentation results were obtained by Bayesian criterion. Experimental results show that the algorithm has strong adaptability, can better overcome the impact of noise on image segmentation. Compared with similar algorithms, the proposed algorithm has higher segmentation accuracy for brain MR images with noise, and obtains better segmentation results.

The existing image segmentation algorithms based on Markov random field are prone to over segmentation and the segmentation results are not ideal. This paper presents an improved image segmentation algorithm based on Markov random field and region merging. First, the algorithm uses the image segmentation algorithm based on the theory of Markov random field and Gaussian mixture model to get the initial segmentation results; second, the region distance between each region is given by using the adjacent relationship, color relationship and boundary condition of each region; finally, the initial segmentation is performed according to the distance between regions and the change rate of color divergence after region merging. The final image segmentation results are output by region merging. In this paper, Berkeley standard image library is used for experimental simulation, and the Dice and Jaccard coefficients are used as the evaluation index of this paper. The experimental simulation shows that the proposed algorithm has better segmentation effect than the existing algorithm based on MRF theory.