给出了一种用于立体图像匹配的改进置信传播算法。基于视差场的连续性假设，传统视差估计置信传播算法将稠密视差场抽象为一种马尔可夫场，置信传播在消息双向传递的马尔可夫网络上进行。考虑到在物体遮挡区域视差场并不连续，首先采用基于初始视差估计的交叉不稳定检测技术检测出遮挡区域，将稠密视差场更加精确地抽象为一种马尔可夫场和贝叶斯场的混合场，置信传播在马尔可夫和贝叶斯的混合网络上进行，使得遮挡区域像素视差信息不传递给非遮挡区域，提高了视差估计精度并降低了算法复杂度。采用Middlebury网站提供的标准测试图像对本文算法进行了客观评估，实验结果表明，本文算法同时具有较好的视差估计精度和运算效率。

This paper proposes an improved belief propagation algorithm for stereo matching.Based on the assumption that the disparity field is continuous,traditional methods regard the disparity field as a Markovian network that transmits two-way information.But in the occluded area,disparity is not continuous.So we propose a new method.Firstly,we use the cross-check technology based on the initial disparity to detect the occluded area.Secondly,we regard the disparity map as a mixed network of Markovian field and Bayes filed.Then the occluded area does not transmit information to the non-occluded area so as to reduce the computational cost of disparity matching.We use the standard test images to evaluate our algorithm.The result shows that the proposed method achieves a high accuracy and efficiency.