面向深度学习密集匹配的无监督损失函数  

Unsupervised loss function for dense matching in deep learning

作　　者：刘潇 官恺 金飞[1] 芮杰[1] 王淑香[1] 林雨准 程传祥 LIU Xiao;GUAN Kai;JIN Fei;RUI Jie;WANG Shuxiang;LIN Yuzhun;CHENG Chuanxiang(Institute of Geospatial Information,Information Engineering University,Zhengzhou 450001,China;The Technical Division of Surveying&Mapping of Xi'an,Xi'an 710054,China)

机构地区：[1]信息工程大学,河南郑州450001 [2]西安测绘总站,陕西西安710054

出　　处：《测绘通报》2024年第11期133-139,共7页Bulletin of Surveying and Mapping

摘　　要：随着深度学习的发展,监督型密集匹配网络取得了瞩目的成就。然而,密集匹配的真实标签制作困难,获取成本高,基于无监督深度学习的密集匹配方法是未来趋势。目前,面向无监督密集匹配提出了众多损失函数,但各种组合损失复杂且存在效果不明的问题。为此,本文针对密集匹配无监督损失函数开展研究,分析各类损失的精度和匹配效果,并考证组合应用的有效性。结果表明,重构相似性约束是令无监督密集匹配网络匹配精度收敛的关键项,组合使用重构损失和左右一致损失有助于无纹理、弱纹理区域匹配,在此基础上加入相对平滑损失能更好适应阴暗环境。With the advancement of deep learning,supervised dense matching networks have achieved remarkable progress.However,obtaining real annotations for dense matching is challenging and costly,making unsupervised deep learning-based methods the future trend.Recently,numerous loss functions have been proposed for unsupervised dense matching.However,their combinations are complex and effects remain unknown.Therefore,this study investigates unsupervised loss functions for dense matching,analyzes the accuracy and matching performance of various losses,and validates the effectiveness of combined applications.The results demonstrate that the appearance matching loss plays a pivotal role in achieving convergence in accuracy for unsupervised dense matching networks.Combining appearance matching loss with left-right disparity consistency loss facilitates accurate non and weak textured region matches.Then,adding relative smoothing loss can better adapt to dark environments.

关 键 词：密集匹配 深度学习 无监督 损失函数 

分 类 号：P237[天文地球—摄影测量与遥感]