摘要
针对同一场景中多组图像匹配精度低,以及不同场景中图像匹配鲁棒性差等问题,提出一种基于深度学习的端到端图像匹配方法。本文基于LoFTR改进,在融合ResNet和FPN的特征提取网络中,减少其激活函数,并加入CBAM注意力模块,后续通过Transfromer生成特征图,进行图像匹配。结果表明,改进网络结构后的方法具有更好的匹配精度以及鲁棒性,在公开数据集Megadepth的子集上进行训练和测试,平均匹配精度达到93.68%,相较于LoFTR提升了1.33%,在阈值为5°、10°、20°情况下姿态误差的AUC数值均有提升;在TUM数据集子集上验证,其平均匹配精度提高了2.13%。
Aiming at the low precision of multiple image matching in the same scene and the poor robustness of image matching in different scenes,an end-to-end image matching algorithm based on depth learning is proposed.Improving on LoFTR,this method reduces its activation functions in the feature extraction network which integrates ResNet and FPN,adds CBAM attention module,and subsequently generates feature maps through Transfromer for image matching.The experiment results show that the improved network structure has better matching precision and robustness.After training and testing on the subset of open dataset Megadepth,the mean matching precision reaches 93.68%,which is 1.33%higher than LoFTR,and the AUC of the pose error at threshold(5°,10°,20°)is also improved;The mean matching precision is improved by 2.13%on TUM dataset subset.
作者
李维文
冉昌艳
LI Weiwen;RAN Changyan(College of Computer and Information Technology,China Three Gorges University Hubei,Yichang,443002,China;Hubei Key Laboratory of Intelligent Visual Monitoring for Hydropower Engineering,China Three Gorges University Yichang,443002;Yichang Key Laboratory oflntelligent Visual Monitoring for Hydropower Engineering China Three Gorges University Yichang,443002;China Three Gorges University,Institute of Electronics and Communications,Yichang,443002)
出处
《长江信息通信》
2023年第2期5-7,11,共4页
Changjiang Information & Communications
基金
2020年宜昌市科学技术研究项目(A20-3-004)
湖北省水电工程智能视觉监测重点实验室开放基金(2020SDSJ07)
水电工程智能视觉监测湖北省重点实验室建设(2019ZYYD007)。
