For the past several years, a variety of algorithms have been focused on how to exploit two-dimensional scanline optimization to augment one-dimensional ones for semi-global matching. Different from the former contributions, an ensemble learning with advanced fast image filtering features for semi-global matching is proposed in this paper. Firstly, fewer categories of features (confidence measures) are extracted through various advanced fast image filters on the original scale of 8 directions’ semi-global matching disparity maps. Then, all the features are weaved together and divided into positive and negative samples for ensemble learning after comparing with ground truth. After that, the initial disparity map is obtained by leveraging the confidence probability of ensemble learning prediction. Finally, an efficient two-step single view disparity refinement strategy is employed, which no longer requires the right view’s disparity map for attaining the final refined results. Performance evaluations on Middlebury v.2 and v.3 stereo data sets demonstrate that the proposed algorithm outperforms other four most recent stereo matching algorithms. In addition, the presented algorithm shows relative high implementation efficiency compared with others.

在过去的几年中，各种算法都集中在如何利用二维扫描线优化来增加一维扫描线以进行半全局匹配方面。与以前的贡献不同，本文提出了一种具有高级快速图像过滤功能的半全局匹配集成学习。首先，在8个方向的半全局匹配视差图的原始比例上，通过各种高级快速图像滤波器提取的特征类别（置信度）较少。然后，将所有特征编织在一起，并与地面实况进行比较，将其分为正样本和负样本进行整体学习。之后，通过利用整体学习预测的置信度获得初始视差图。最后，采用了一种有效的两步单视图视差细化策略，该策略不再需要正确的视图视差图来获得最终的细化结果。米德尔伯里绩效评估v.2和v.3立体声数据集证明，该算法优于其他四种最新的立体声匹配算法。另外，与其他算法相比，该算法具有较高的实现效率。