Research On Local Stereo Matching Algorithm For Deep Image

Computer vision is a cross field with artificial intelligence, robotics, machine learning, geometry and signal processing. With the rapid development of computer technology, the application field of computer vision is continuously expanded. Using computer vision, people obtain the 3D information from 2D images, which makes stereo vision be a hot field in computer vision. Stereo vision is an important research field of computer vision, from two or more images captured by camera or camera array, the purpose of the study is to find the corresponding pixel in order to recover the 3D scene information and cognitive 3D world. Stereo vision includes three processes: camera calibration, stereo matching and 3D reconstruction. The realization of 3D reconstruction is obtained by using the corresponding relationship of same point in different images, camera parameter matrix and triangular geometric relationship. Obviously, the stereo matching is a hot research, and its precision directly affects the performance of 3D reconstruction. Stereo matching is a hot and difficult topic on computer vision because of the occlusion problem, depth discontinuities, textureless region or texture of monotonous repetition. In this paper, based on the related theories of further study of stereo matching algorithm, the main work and innovations are as follows:1. A local stereo matching algorithm is modified based on adaptive support construction. This design is mainly use the image segmentation technology to re-define the supporting region which will also be used on the post-processing. First of all, the image segments using mean-shift segmentation, using the results, each anchor pixel gets a quadruple to denote the coarse left, right, up and bottom arm length, then an initial support skeleton is defined by the quadruple as well as the color similarity. Secondly, cost aggregation is calculated by combining truncation brightness difference and gradient. Finally, after obtaining the initial disparity map using WTA, the disparity map is proposed with adaptive weight median filter. The experimental results show that this scheme achieves better performance than many local stereo matching algorithms.2. A new local stereo matching algorithm is modified based on adaptive weight on the boundary pixels. Since the segmentation effect the accurate of construction region, using Gestalt grouping, we take account the weight of each surrounding pixels of regions. Similar to the third chapter, firstly, a quadruple is decided by the segmentation, then re-construct each region using the correlation between adjacent pixels and quadruple. Then, for the most outer pixels of each region, we determine the weight by Gestalt criteria. For cost aggregation, we adopt the same method of third chapter. Finally, after obtaining the initial disparity map, iterative region voting is proposed to post-process the disparity map. From the experimental result, this method achieves the desirable effect.