Research on stereo Matching algorithm based on neural network

In order to achieve high precision stereo matching in the case of limited computing resources, an end-to-end parallax computing network is proposed by improving DispNet network. The network uses code-decoding structure, adds residual structure and cost calculation method based on matrix multiplication to optimize the network structure, and adds ECA module to the parallax estimation network for end-to-end training, adaptively enhances the important channel information in the matching cost body, and improves the learning ability of the network. The experimental results show that the improved parallax computing network can achieve dense and high-precision parallax prediction in real indoor and outdoor scenes, with good generalization ability, and significantly improve the prediction accuracy of weak texture, fine structure and other ill regions, which fully proves the robustness and effectiveness of the proposed model.