现有的多尺度立体匹配算法对各尺度的代价函数采用相同权值,而忽略了各尺度层对整个匹配代价的不同影响,增加了误匹配点。针对此问题,提出了自适应权值的跨尺度立体匹配算法框架。采用统一的代价聚合函数框架在不同尺度上进行代价匹配,并提出利用各像素窗口的信息熵作为不同尺度下匹配代价对整个匹配代价的影响因子;同时为了保证不同尺度下同一像素的代价一致性,在代价函数里加入正则化因子。本文算法框架可以应用在利用多尺度进行代价匹配的算法上,并使原有算法的准确率和稳健性得到提高。基于本文算法框架,分别采用不同代价聚合函数在Middlebury数据集上进行测试。为保证测试的公平性,各算法均未进行后续的视差求精步骤,实验表明,本文算法有效地提高了多尺度立体匹配的准确率和稳健性。

In the existing weight-adaptive cross-scale algorithms, the same weight for the cost function of each scale is adopted, the different influence of each scale layer on the whole matching cost is missing, and thus the number of mismatching points increases. As for this problem, a weight-adaptive cross-scale algorithm framework for stereo matching is proposed. The cost matching is performed on different scales in the framework of unified cost aggregation function and the information entropy of each pixel window is used as the influence factor of the matching cost at each scale on the whole matching cost. At the same time, a regularization factor is added to the cost function to ensure the cost consistency at different scales for the same pixel. The proposed algorithm framework can be applied to the multi-scale algorithm of cost matching and improve the accuracy and robustness of the existing algorithms. Based on the proposed algorithm framework, the different cost aggregate functions are tested on the Middlebury dataset. To ensure the fairness of tests, as for each algorithm, there is no a subsequent parallax refinement step. The experimental results show that the proposed algorithm effectively improves the accuracy and robustness of multi-scale stereo matching.