Abstract: Based on PSMNet stereo matching network，an improved stereo matching algorithm with shallow struc-ture and wide receptive field -SWNet was proposed，in order to solve the stereo matching problem in binocular vi-sion，reduce the number of parameters of the stereo matching network，reduce the computational complexity of the algorithm，and improve the practicability of the algorithm． The shallow structure means fewer layers，fewer pa-rameters and faster processing speed，while wide receptive field means that the network is more receptive and can acquire and retain more spatial information． SWNet consists of three parts: feature extraction，3D convolution and disparity regression． In the aspect of feature extraction，Atrous Spatial Pyramid Pool (ASPP) was introduced，which was used to extract multi-scale feature information． Feature fusion module was designed to fuse multi-scale feature information and build matching cost volume． The 3D convolutional neural network use the stack encoding
and decoding structure to further regularize the matching cost volume and obtain the corresponding relationship be-tween the feature points under different disparity conditions． Finally，the disparity map was obtained by regres-sion． SWNet performed well on both SceneFlow and KITTI 2015 public datasets，with a 48. 9% reduction in the number of parameters and a 2. 24% mismatching rate compared to the reference algorithm PSMNet．

Key words: stereo matching, PSMNet stereo matching network, convolutional neural network, deep learning, At-rous Spatial Pyramid Pooling, spatial feature information, feature fusion module