| 3D reconstruction is widely used in the field of computer vision.Stereo matching is not only a crucial step in 3D reconstruction,but also the most difficult step.Therefore,the research on end-to-end stereo matching network can realize the overall optimization of 3D reconstruction system.When studying the current advanced end-to-end stereo matching network,this paper puts forward the idea of deep integration of traditional algorithm and deep learning model.On the one hand,the real prediction value of one-dimensional Semi-Gobal Stereo Matching is added when the left and right views are input,so that the network converges quickly;On the other hand,a new convolution input is added to enable the model to directly learn the real features of Semi-Gobal Stereo Matching.In terms of model structure optimization,this paper uses the characteristics of automatic learning of weighted information by attention mechanism,combined with the construction idea of previous advanced end-to-end stereo matching network,proposes an end-to-end stereo matching network sgma net based on Semi-Gobal Stereo Matching feature extraction and attention mechanism,and designs a series of ablation experiments to obtain the optimal structure of the network.By comparing with the results of the baseline model,It is verified that sgma net improves the convergence speed and prediction accuracy,especially in the edge and non textured areas.The main content and contributions of this paper are as follows:1)Research on several advanced end-to-end stereo matching networks,and analyze the ideas of end-to-end network construction and optimization,which is the basis of designing and improving end-to-end stereo matching networks.It is found that there is room for improvement and Optimization in feature extraction and matching cost volume construction.2)The real estimated parallax of Semi-Gobal Stereo Matching based on OpenCV is introduced,and the traditional stereo matching algorithm is deeply integrated with the end-to-end network.The specific structure of the improved end-to-end stereo matching network is designed through experiments,and the loss function of network training is optimized.Experiments show that the improved network improves the prediction accuracy.3)Based on the improved network,the attention mechanism is further optimized,and an end-to-end stereo matching network sgma net based on Semi-Gobal Stereo Matching feature extraction and attention mechanism is proposed.Through a series of ablation experiments,the specific structure of the attention mechanism is designed.At the same time,it is compared with other models,and it is found that sgma net is the best model relative to the baseline,The parallax prediction error is relatively reduced by 18.7%,and the number of convergence rounds of the model is reduced by about 30%. |
