基于全连接神经网络和边缘感知视差传播的立体匹配算法研究  

Stereo matching algorithm based on fully-connected neural network and edge-aware disparity propagation

作　　者：严邓涛 霍智勇[2] 戴伟达 陈钊 YAN Dengtao;HUO Zhiyong;DAI Weida;CHEN Zhao(College of Telecommunications ＆ Information Engineering, Nanjing University of Posts and Telecommunications, Nanjing 210003, China;Jiangsu Province Key Lab of Image Processing and Image Communication, Nanjing University of Posts and Telecommunications, Nanjing 210003, China)

机构地区：[1]南京邮电大学通信与信息工程学院,江苏南京210003 [2]南京邮电大学江苏省图像处理与图像通信重点实验室,江苏南京210003

出　　处：《南京邮电大学学报（自然科学版）》2018年第3期83-88,共6页Journal of Nanjing University of Posts and Telecommunications：Natural Science Edition

基　　金：国家自然科学基金(61471201);江苏省高校自然科学研究重点项目(13KJA510004);江苏省自然科学基金青年基金(BK20130867);江苏省"六大人才高峰"项目(2014-DZXX-008);南京邮电大学"1311人才计划"资助项目

摘　　要：提出一种基于全连接神经网络和边缘感知视差传播(Edge-aware Disparity Propagation,EDP)的立体匹配算法。首先通过全连接网络获得初始的匹配代价,并利用胜者为王(Winner Takes All,WTA)算法生成初始视差图;然后进行一致性检测,对不一致视差点采用EDP和测地距滤波算法计算得到新的匹配代价和视差值,并填充初始视差图;最后,对填充后的视差图进行亚像素增强优化。实验数值表明,该算法能有效降低立体匹配的误匹配率,提高视差图精度,特别是在非闭塞区域有较低的误匹配率。A stereo matching algorithm based on fully connected neural network and edge-aware disparity propagation （EDP） is presented. Firstly, a fully-connected neural network is used to obtain the raw matc- hing cost and the raw disparity map by winner-take-all （WTA） algorithm;Then, the inconsistent pixels are filtered out by the consistency detection, the inconsistent disparity point is calculated by using EDP and geodesic distance filtering algorithm to obtain new matching cost and disparity value, and the initial disparity map is filled. Finally, the sub-pixel enhancement optimization is performed on the filled disparity map. The experimental results show that the algorithm can effectively reduce the mismatch rate of stereo matching and improve the accuracy of the disparity map, especially in the non-occluded area with a low false matching rate.

关 键 词：立体匹配 全连接神经网络 视差传播 测地距滤波 边缘感知 亚像素增强 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]