Efficient Stereo Matching Based on Depthwise Separable Convolution and Attention Mechanisms

Currently, conventional deep learning stereo matching networks can produce high-precision disparity maps, but they suffer from issues such as large parameter counts and slow processing speeds. Addressing these issues, a lightweight stereo matching network structure is designed by combining depthwise separable convolutions and attention mechanisms in this paper. This approach ensures that the disparity maps obtained by the network have high accuracy, while reducing the network's parameter count and improving its processing speed. Compared to the high-precision network ACVNet, our network has an end-point error metric that is 0.16px higher, but its parameter count is 55.4% lower, and its processing speed is 80.5% faster than ACVNet.