多成本融合的立体匹配网络  

Stereo Matching Network with Multi-Cost Fusion

作　　者：张锡英[1] 王厚博 边继龙[1] ZHANG Xiying;WANG Houbo;BIAN Jilong(School of Information and Computer Engineering,Northeast Forestry University,Harbin 150040,China)

机构地区：[1]东北林业大学信息与计算机工程学院,哈尔滨150040

出　　处：《计算机工程》2022年第2期186-193,共8页Computer Engineering

基　　金：黑龙江省自然科学基金(F2018002)。

摘　　要：立体匹配网络中的特征提取是提高双目视觉立体匹配精确度的关键步骤。为充分提取图像特征信息,结合密集空洞卷积、空间金字塔池化和堆叠沙漏的特点,构建一种多成本融合的立体匹配网络DCNet。引入密集空洞卷积和空间金字塔池化方法提取多尺度特征信息,同时使用轻量化注意力模块优化多尺度特征信息,构建多特征融合的匹配代价卷。在此基础上,利用3D卷积神经网络和堆叠沙漏网络聚合匹配代价信息,并通过回归的方式生成视差图。实验结果表明,该网络在KITTI2015数据集上的误匹配率为2.12%,相比PSMNet、DisNetC、PDSNet等网络,在特征提取部分能够获得更丰富的特征信息,且提升特征匹配的效果。In stereo matching networks,the feature extraction process is key for improving the accuracy of binocular stereo matching.To extract image feature information,this study combines the characteristics of dense atrous convolution,spatial pyramid pooling,and stacked hourglass to propose stereo matching network DCNet with multi-cost fusion.The network introduces dense atrous convolution and spatial pyramid pooling to extract multi-scale feature information.The lightweight attention module is used to optimize the multi-scale feature information as well as construct the matching cost volume of multi feature fusion.The matching cost information is aggregated by the 3 D convolution neural network and stacked hourglass network,whereby the parallax map is generated by regression.The experimental results on KITTI2015 data set show that the false matching rate of this network is 2.12%.Compared with PSMNet,DISNetC,and PDSNet,the proposed method provides richer feature information in feature extraction and improves the effect of feature matching.

关 键 词：立体匹配 密集神经网络 深度卷积神经网络 深度学习 注意力机制 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]