High-Resolution Stereo Matching Network Based on Attention Mechanism and Intelligent Computing

To solve the problem that the stereo matching algorithm loses the highest resolution feature level information in the process of feature extraction, a high resolution unit is used to maintain the highest level feature information and fuse the feature information of each resolution level. Due to the lack of semantic correlation and the measurement of the relationship between each cost feature in convolution process, the mismatch rate, especially in the foreground region, is too high. This paper uses the attention mechanism module which combines channel attention mechanism and spatial attention mechanism. By comparing Scene Flow and KITTI datasets, the proposed algorithm is compared with GC-net and other commonly used algorithms. On KITTI datasets, the proposed algorithm reduces the false matching rate by 0.26 percentage points compared with GC-net, and reduces the false matching rate by 0.1 and 1.5 percentage points in background and foreground regions respectively.