The Research Of Stereo Matching Based On Deep Learning

Stereo matching is one of the important problems in computer vision,the technology reconstructs scenes through different perspectives of multiple images.The key point in the corresponding stereo matching(that is,the acquisition of disparity map)is to find the same point from a different perspective at in the scene.However,in the presence of reality scene,stereo matching has numerous challenges:reflection,shelter,disparity discontinuities,perspective distortion,and other non-texture area.So in recent years taken from real-life scenarios precision disparity map has been the focus of attention of domestic and foreign researchers.Stereo matching generally involves four steps:matching cost computation,cost aggregation,disparity optimization and disparity refinement.Local stereo matching algorithm disparity calculation depends only on the value of the local window,usually carried out by the implicit assumption of smoothing assumptions.Cost aggregation is usually skipped by the global stereo matching algorithm,performed an obvious disparity smoothing hypothesis,and then solve an optimization problem.Deep learning in recent years has been focused closely in academia and industry,through multiple nonlinear transformation,learning automatically large amounts of data from multi-feature,replacing traditional machine learning manual design features,deep structure with strong learning ability and communication skills,good at extracting complex global features and context information.Wherein the convolution neural network has been widely used in vision problems.General stereo matching algorithms could not solve multiple problems at the same time.This paper presents a stereo matching based on a full convolution neural network,perform end-to-end supervised learning.Full convolution neural network can perform pixel level prediction,with a lot of ground truth disparity map acquired by laser scanning,and then use the full convolution neural network image of different perspectives to the disparity map mappings.Final results show that this method can solve stereo matching problem.