作者投稿和查稿 主编审稿 专家审稿 编委审稿 远程编辑

计算机工程 ›› 2022, Vol. 48 ›› Issue (2): 186-193. doi: 10.19678/j.issn.1000-3428.0062144

• 图形图像处理 • 上一篇    下一篇

多成本融合的立体匹配网络

张锡英, 王厚博, 边继龙   

  1. 东北林业大学 信息与计算机工程学院, 哈尔滨 150040
  • 收稿日期:2021-07-21 修回日期:2021-08-21 发布日期:2021-09-08
  • 作者简介:张锡英(1971-),女,副教授,主研方向为图像处理、数据挖掘;王厚博,硕士研究生;边继龙(通信作者),讲师、博士。
  • 基金资助:
    黑龙江省自然科学基金(F2018002)。

Stereo Matching Network with Multi-Cost Fusion

ZHANG Xiying, WANG Houbo, BIAN Jilong   

  1. School of Information and Computer Engineering, Northeast Forestry University, Harbin 150040, China
  • Received:2021-07-21 Revised:2021-08-21 Published:2021-09-08

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。

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

Abstract: 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 3D 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.

Key words: stereo matching, dense neural network, deep Convolutional Neural Network(CNN), Deep Learning(DL), attention mechanism

中图分类号: