Research Of Stereo Matching Based On Deep Learning

As a basic and important research content in computer vision,the essence of stereo matching is to calculate the disparity value of each pixel in the target image based on a pair of recitfied images.It is widely used in 3D ranging,robot navigation and automatic driving.With the continuous development of deep learning,the accuracy of stereo matching algorithms based on deep learning farly exceeds the accuracy of traditional stereo matching algorithms.However,the current methods still exist the problems of low matching accuracy in complex regions such as textureless area,reflective surfaces,and occlude area.Using the target's multi-scale and contextual information effectively can solve the above problems.Therefore,the research focus of this article is how to effectively apply multi-scale and contextual information to each stage of stereo matching to improve the accuracy of algorithms and models degree.The main research contents of this article are as follows:(1)In terms of cost computation,a cross-scale cost computation method that mimics human visual mechanism is proposed.The approximate depth of the pixel is obtained from multi-scale feature information and it can effectively guides the subsequent cost aggregation process.Combining with the cost computation method in the form of connection can achieve more great results.Experimental results show that the proposed method achieved better results compared with existing methods.(2)In terms of feature extraction and feature processing,a hybrid pyramid pooling module and a multi-scale 3D feature fusion module are proposed.The hybrid pyramid pooling module combines the characteristics of both spatial pyramid pooling module and atrous pooling module,which can extract richer context information,and multi-scale 3D feature fusion module can effectively process the feature information in the cost volume.(3)In terms of cost aggregation,a learning-based cost aggregation algorithm is proposed.It is implemented by two sub-networks,the aggregation sub-network and the guidance sub-network.The aggregation sub-network is used to aggregate cost values from different dimensions,and the guidance sub-network is used to extract structural information from reference image as global view guidance.The experiment results show that the proposed algorithm and sub-network can be embedded into existing stereo networks in an end-to-end manner and improve performance of the model.