基于ConvNeXt和可变形卷积的多视图重建  

Multi-view stereo reconstruction based on ConvNeXt and deformable convolution

作　　者：刘韵婷 徐利钦 LIU Yunting;XU Liqin(School of Automation and Electrical Engineering,Shenyang Ligong University,Shenyang 110159,China)

机构地区：[1]沈阳理工大学自动化与电气工程学院,辽宁沈阳110159

出　　处：《通信与信息技术》2024年第6期35-38,43,共5页Communication & Information Technology

摘　　要：三维重建技术是目前计算机视觉领域的热点研究项目,目的是通过图像或者图像集来还原物体的几何形状。针对多视图立体(Multi-view Stereo,MVS)重建结果整体以及重建完整度不理想的问题,对特征提取模块和损失函数进行优化,提出了一种基于ConvNeXt和可变形卷积的多视图立体重建网络;该网络在DTU数据集上进行训练和测试,与CasMVSNet相比,重建的完整度指标提高了24.2%,重建整体性指标提高了2.5%,比其他的传统方法和部分现有的深度学习方法完整度或整体都有提升;对DTU数据集测试的结果进行分析表明,该网络框架在整体性能上有着出色的结果,可以获得质量更优的三维重建效果。In the field of computer vision,3D reconstruction technology is a hot research topic,which aims to restore the geometric shape of objects through images or image sets.In order to solve the problem of unsatisfactory overall and completeness of multi-view stereo(MVS)reconstruction results,the multi-view stereo reconstruction network based on ConvNeXt and Deformable convolution was proposed by optimizing the feature extraction module and loss function.After training and testing on DTU data set,compared with CasMVSNet,the index of completeness of reconstruction increased by 24.2%and the index of completeness of reconstruction increased by 2.5%,which is better,compared with other traditional methods and many existing methods of deep learning.The test results of DTU data set show that the network has good performance in the overall performance,and can obtain better quality 3D reconstruction effect.

关 键 词：可变形卷积 多视图立体(MVS) 三维重建 ConvNeXt 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]