欢迎访问《图学学报》 分享到：
图学学报

图学学报 ›› 2022, Vol. 43 ›› Issue (5): 875-883.DOI: 10.11996/JG.j.2095-302X.2022050875

• 图像处理与计算机视觉 • 上一篇    下一篇

基于多尺度特征递归卷积的稠密点云重建网络 

  

  1. 长安大学信息工程学院，陕西 西安 710064
  • 出版日期:2022-10-31 发布日期:2022-10-28
  • 基金资助:
    国家自然科学基金面上项目(61771075)；陕西省自然科学基金项目(2017JQ6048)；广西精密导航技术与应用重点实验室项目(DH201711) 

Dense point cloud reconstruction network using multi-scale feature recursive convolution

  1. School of Information Engineering, Chang’an University, Xi’an Shaanxi 710064, China
  • Online:2022-10-31 Published:2022-10-28
  • Supported by:
    National Natural Science Foundation of China (61771075); Natural Science Foundation of Shaanxi Province (2017JQ6048); Guangxi Key Laboratory of Precision Navigation Technology and Application, Guilin University of Electronic Technology (DH201711)

PDF (PC)

83

摘要：

针对在三维重建任务中，由于弱纹理区域的光度一致性测量误差较大，使得传统的多视图立体 算法难以处理的问题，提出了一种多尺度特征聚合的递归卷积网络(MARDC-MVSNet)，用于弱纹理区域的稠 密点云重建。为了使输入图像分辨率更高，该方法使用一个轻量级的多尺度聚合模块自适应地提取图像特征， 以解决弱纹理甚至无纹理区域的问题。在代价体正则化方面，采用具有递归结构的分层处理网络代替传统的三 维卷积神经网络(CNN)，极大程度地降低了显存占用，同时实现高分辨率重建。在网络的末端添加一个深度残 差网络模块，以原始图像为指导对正则化网络生成的初始深度图进行优化，使深度图表述更准确。实验结果表 明，在 DTU 数据集上取得了优异的结果，该网络在拥有较高深度图估计精度的同时还节约了硬件资源，且能 扩展到航拍影像的实际工程之中。

关键词: 深度学习, 计算机视觉, 遥感测绘, 三维重建, 多视图立体, 递归神经网络

Abstract:

In the task of 3D reconstruction, it is difficult to deal with the traditional multi view stereo algorithm because of the large photometric consistency measurement error in the weak texture region. To solve this problem, a recursive convolution network of multi-scale feature aggregation was proposed, named MARDC-MVSNet (multi-scale aggregation recursive multi view stereo net with dynamic consistency), which was utilized for dense point cloud reconstruction in weak texture areas. In order to boost the resolution of the input image, this method used a lightweight multi-scale aggregation module to adaptively extract image features, thereby addressing the problem of weak texture or even no texture region. In terms of cost volume regularization, a hierarchical processing network with recursive structure was used to replace the traditional 3D CNN (convolutional neural networks), greatly reducing the occupation of video memory and realizing high-resolution reconstruction at the same time. A depth residual network module was added at the end of the network to optimize the initial depth map generated by the regularized network under the guidance of the original image, so as to produce more accurate expressions of the depth map. The experimental results show that excellent results were achieved on the DTU data set. The proposed network can not only achieve high accuracy in depth map estimation, but also save hardware resources, and it can be extended to aerial images for practical engineering. 

Key words:  , deep learning, computer vision, remote sensing mapping, 3D reconstruction, multi view stereo, recurrent neural network 

中图分类号: