基于信任度传播的体视算法

A Belief Propagation Algorithm for Stereo Matching

针对信任度传播算法计算量大及误匹配率高的问题,提出一种高效的计算稠密视差图的全局优化算法.首先,根据像素匹配代价的特点、视差不连续亮度变化的特征,定义具有适应性的数据约束和平滑约束,并对平滑约束进行分层调节后执行消息的传输.其次,讨论消息传输迭代过程中的冗余计算问题,通过检测消息的收敛性减少运行时间.最后,分析信任度传播算法中的误匹配问题,通过匹配的对称性检测遮挡,并提出重建数据项后,利用贪婪迭代法优化所得视差图,将图像中可靠像素的视差向不可靠像素扩散.实验结果表明,该算法能以较快的速度计算出更理想的视差图.

Large-scale computing and high matching error rate are two disadvantages of the existing algorithms based on belief propagation. An efficient global optimal algorithm for dense disparity mapping is presented. Firstly, according to the feature of match cost and the disparity discontinuities accompanying intensity changes, the adapted data constraint and the smoothness constraint are defined, and the messages are passed after the smoothness constraint is adjusted in every level, Then, the redundancy in message passing iteration process is discussed, and the message convergence is checked to decrease the running time. Finally, match symmetry is used to detect occlusions according to the analysis of matching errors in belief propagation algorithm. After the data term is reconstructed, a greedy method is used to iteratively refine the disparity result to propagate disparity information from the stable pixels tO the unstable ones. The experimental results show the proposed algorithm computes a disparity map accurately with relative less time.