针对立体匹配算法在图像非遮挡区域,尤其是弱纹理区域匹配精度较低的问题,提出一种基于边缘约束迭代的非局部立体匹配算法。该算法结合颜色和梯度信息构建匹配代价计算函数;根据左右目图像分别构建最小生成树,结合图像平滑度对代价函数值进行代价聚集,并对赢者通吃策略得到的视差图进行边缘检测,将图像边缘作为约束性条件对代价值再次进行代价聚集以优化结果;最后通过视差求精得到稠密的视差图。实验结果表明,在Middlebury测试平台上该算法对31组标准图像对中非遮挡区域的平均误匹配率为8.35%;与其他5种方法比较可知,本文算法有效提高了非遮挡区域匹配的准确度。

To address the problem of stereo matching algorithms having low matching accuracy in non-occluded regions, especially in the weak-textured regions, a non-local stereo matching algorithm based on edge constraint iteration was proposed. Firstly, the proposed method combined the color and gradient information to construct a matching cost computation function. Secondly, the minimum spanning tree structures of left and right images were constructed, and the cost volumes were aggregated according to the smoothness information of the image. Subsequently, a disparity map obtained by the winner-takes-all strategy was used for edge detection. The image edges were then used as constraints to re-aggregate the cost volumes and optimize the results. Finally, dense disparity maps were obtained by the disparity refinement process. The experimental results demonstrate that, for 31 pairs of images from the Middlebury test platform, the average mismatching rate in non-occluded regions of the proposed algorithm is 8.35%. Compared with five existing methods, the proposed algorithm can effectively improve matching accuracy in non-occluded regions.