一种基于最小割的稠密视差图恢复算法

An Algorithm for Dense Disparity Mapping by Using Graph Cuts

针对经典最小割算法计算量大和适应性不足的问题,提出一种改进的基于网络最小割计算稠密深度图的全局优化方法.首先,根据视差变化与不连续区域之间的关系,定义了具有一定适应性的平滑约束和遮挡约束,然后使用网络最小割算法,求解遮挡情况下的稠密视差.其次,在分析最小割算法复杂性的基础上,给出了一种受限α-扩展(α-expansion)操作,该操作根据灰度连通性和特征点匹配的结果对每次网络构造的顶点进行控制,减少网络中顶点和边的数目,可有效提高计算效率.实验结果显示,该算法在保证视差恢复准确性的前提下,能以较快的速度计算出较理想的稠密视差图.

Vast computation is a great disadvantage of the existing graph cuts based vision algorithms. Lack of adaptability is another issue. An improved global optimal algorithm for dense disparity mapping using graph cuts is presented in this paper. First, adapted occlusion penalty and smoothness penalty are defined based on the intrinsic relation between the disparity changes and the discontinuities in an image. The graph cuts based algorithm is employed to get an optimial dense disparity mapping with occlusions. Secondly, according to the complexity analysis of graph cut algorithms, an operation named restricted α-expansion operation is defined to control the vertexes generation during graph constructing based on the result of normalized correlation algorithm. It is a great help to reduce the vertexes and edges in the constructed graph, thus the computing is speeded up. The experimental results show performance of the proposed algorithm is improved and it will take a shorter time to compute an accuracy dense disparity mapping.