Stereo Matching Algorithm Based on DACVNet

Stereo matching is the most important step in binocular vision 3D reconstruction, and the matching effect directly affects the result of 3D reconstruction. In this paper, ACVNet is used as the benchmark network, and the network model is improved. The reference network uses the three-layer ResNet structure in the image classification task in the process of feature extraction, and there are some problems such as insufficient edge and texture information extraction and inaccurate feature information extraction. To solve these problems, based on ACVNet, a lightweight DFA module is introduced, which is more suitable for feature extraction in stereoscopic matching tasks. Considering the poor effect of the fine texture region, an inter-dimensional multi-scale module is introduced into the construction of the matching cost, and the feature map obtained after feature extraction is processed to enhance the effect of the weak texture region. The parallax completion module and parallel multi-scale module are introduced into the cost aggregation to reduce the errors in the occlusion region and obtain a denser high-precision parallax map. The results of training and evaluation on SceneFlow, KITTI2012 and KITTI2015 data sets show that the accuracy of the proposed improved algorithm is improved compared with the benchmark model. Finally, the three-dimensional matching method is used to reconstruct the real scene, and good results are obtained.