基于狄利克雷分布和参数分析的高斯混合模型图像分割算法

传统的高斯混合模型对于含有噪声的图像不能进行有效的分割。针对有噪声图像的分割问题，提出了一种基于狄利克雷分布和参数分析的高斯混合模型图像分割算法。首先采用高斯函数对像素计算先验概率值，然后采用狄利克雷分布和定律关联像素间的邻域信息，并利用梯度下降法优化参数。实验结果表明，本文算法对无噪声和有噪声图像的分割结果比传统方法更有效，误分率更低。

Gaussian Mixture Model Image Segmentation Method Based on Dirichlet Distribution and Parameter Analysis

With the development of image processing field, more and more effective image segmentation algorithms, in medical image processing, brain and blood vessel magnetic resonance image segmentation, lung nodules automatic detection and tumors detection have been recently proposed. This paper will consider the brain MRI segmentation problems. It is worthy of noting that although many traditional clustering algorithms of medical image segmentation like K-means, Fuzzy C-means (FCM) and general Gaussian mixture model can effectively deal with these medical images without noise, they cannot perform well in noisy medical images which often exist in original medical image processing. Therefore, it is quite necessary and actually required for researching the image segmentation algorithm with noise elimination. To this end, this paper propses an image segmentation method based on Gaussian mixture model with Dirichlet distribution and parameter analysis, which can effectively denoise the medical images. Firstly, Gaussian function is adopted to calculate pixels'' prior probabilities. Dirichlet distribution and Dirichlet law are utilized to construct the local spatial information among pixels. And then, the gradient descent method is used to optimize the loss function and update model parameters at the same time. Finally, the probability of each pixel for each label can be obtained and the pixel''s highest prediction probability in all labels can be selected as the clustering result. This paper performs four clustering segmentation algorithms on normal creatures and brain MR images with Gaussian noise and salt &pepper noise. The experiment results show that image segmentation on normal creatures with three and four labels have a good result to cluster pixels which especially belong to background. Also, the proposed algorithm used to segment brain MR images with Gaussian noise and salt &pepper noise is more precisely and clearer than other three traditional algorithms. In conclusion, the experiment results show that the proposed method is better than the traditional image segmentation methods which has the lowest segmentation error rate and fine definition.