Research On Binocular Stereo Matching Method Based On Deep Learning

Computer vision is a subject devoted to making computers recognize three-dimensional environmental information through two-dimensional images.Among them,binocular stereo vision technology is favored by many scholars because of its universality,flexibility and low cost.It is widely used in cutting-edge research such as autonomous driving and 3D reconstruction.Stereo matching is the most important link in binocular stereo vision.Due to the difficulty of matching the "morbid areas" such as occlusion,weak texture or repeated texture,there is a problem that the output of the depth information is not accurate due to the poor matching effect.It is the focus and difficulty of scholars' research.In recent years,with the excellent performance of deep learning in target tracking,semantic recognition,image classification and other applications,using deep learning to solve the problem of stereo matching in binocular vision has become a new research direction.The main research work of this thesis is as follows:Aiming at the two problems of the slow convergence rate of the neural network and the network oscillation without convergence in the deep learning method,this thesis conducted a study on the convergence rate of the network training based on the learning rate algorithm.Based on the learning rate annealing algorithm,this thesis improves a learning rate setting method that combines segmentation attenuation and periodic pulse perturbation.Simulation and experimental data show that for ethnic clothing classification,a relatively constant learning rate,warm-up restart learning rate and The exponential decay learning rate algorithm and the improved learning rate algorithm reduce the number of iterations from 317,159,104 to 80 when the network accuracy reaches 80%,which significantly improves the convergence speed.In view of the problem that the left and right images of the "morbid area" often have multiple corresponding points or no corresponding points,this thesis conducted a study of the binocular stereo matching method based on deep learning.By building an end-to-end 3D convolutional neural network,the context information was fully utilized,output high-quality depth map through parallax regression.Simulation and experimental data show that the improved stereo matching algorithm based on deep learning in the Middlebury dataset has a disparity estimation error rate of only 7.74%,which is comparable to five classic methods(Dense-CNN,SGM,MC-CNN,i Res Net,PSMNet),Compared with 0.77%,13.09%,0.9%,17.78% and 36.97%.And it can meet the real-time requirements,while outputting accurate depth images while fully retaining image detail features.