• 中国计算机学会会刊
  • 中国科技核心期刊
  • 中文核心期刊

计算机工程与科学

• 图形与图像 • 上一篇    下一篇

基于像素注意力的双通道立体匹配网络

桑海伟1,3,徐孩2,熊伟程1,左羽1,赵勇1,2   

  1. (1.贵州师范学院数学与大数据学院,贵州 贵阳 550018;2.北京大学深圳研究生院信息工程学院,广东 深圳 518055;
    3.贵州大学计算机科学与技术学院,贵州 贵阳 550025)
  • 收稿日期:2019-09-02 修回日期:2019-11-01 出版日期:2020-05-25 发布日期:2020-05-25
  • 基金资助:

    国家自然科学基金(61771321);贵州省科技计划项目(黔科合支撑[2019]239号,黔科合基础[2019]1250号);贵州省第四批千人创新创业人才;深圳科技计划科研布局项目(JCYJ201605066172651253)

Pixel attention based siamese convolution
neural network for stereo matching

SANG Hai-wei1,3,XU Hai2,XIONG Wei-cheng1,ZUO Yu1,ZHAO Yong1,2   

  1. (1.School of Mathematics and Big Data,Guizhou Education University,Guiyang 550018;
    2.School of Electronic and Computer Engineering,Shenzhen Graduate School,Peking University,Shenzhen 518055;
    3.College of Computer Science and Technology,Guizhou University,Guiyang 550025,China)

     
  • Received:2019-09-02 Revised:2019-11-01 Online:2020-05-25 Published:2020-05-25

摘要:

针对现有立体匹配算法在弱纹理、重复纹理、反射表面等病态区域误匹配率高的问题,提出一种基于像素注意力的双通道立体匹配卷积神经网络PASNet,该网络包括双通道注意力沙漏型子网络和注意力U型子网络。首先,通过双通道注意力沙漏型子网络提取输入图像的特征图;其次,通过关联层得到特征图的代价矩阵;最后,利用注意力U型子网络对代价矩阵进行代价聚合,输出视差图。在KITTI数据集上的实验结果表明,所提出的网络能有效解决病态区域误匹配率高等问题,提升立体匹配精度。
 

关键词: 立体匹配, 像素注意力, 沙漏型子网络, U型子网络, 双通道

Abstract:

Aiming at the problem that the existing stereo matching algorithm has high mismatch rate in ill-posed regions such as weak texture, repeated texture and reflective surface, a new pixel attention siamese neural network is proposed. Our method consists of siamese attention hourglass subnetwork and attention U-shaped subnetwork . Firstly, the feature map of the input image is extracted by the siamese attention hourglass subnetwork. Secondly, the cost matrix of the feature graph is obtained through the correlation layer. Finally, the cost matrix is aggregated by the attention U-shaped subnetwork, and the disparity map is output. Experiments on the KITTI dataset demonstrate that the proposed algorithm can effectively solve the ill-posed problem and improve the stereo matching accuracy.
 

Key words: stereo matching, pixel attention, hourglass subnet, U-shaped subnet, siamese