深度学习技术在影像密集匹配方面的进展与应用  

Progress and Application of Deep Learning in Image Dense Matching

作　　者：张卡 宿东[2,3] 王蓬勃 陈辉 张珊 叶龙杰[2,3] 赵娜 ZHANG Ka;SU Dong;WANG Peng-bo;CHEN Hui;ZHANG Shan;YE Long-jie;ZHAO Na(Key Laboratory of Urban Land Resources Monitoring and Simulation, Ministry of Natural Resources, Shenzhen 518034, China;School of Geography, Nanjing Normal University, Nanjing 210023, China;Key Laboratory of Virtual Geographic Environment, Ministry of Education, Nanjing Normal University, Nanjing 210023, China;Jiangsu Center for Collaborative Innovation in Geographical Information Resource Development and Application, Nanjing 210023, China;Jiangsu Province State Key Laboratory Cultivation Base of Geographical Environment Evolution, Nanjing 210023, China)

机构地区：[1]自然资源部城市国土资源监测与仿真重点实验室,深圳518034 [2]南京师范大学地理科学学院,南京210023 [3]南京师范大学虚拟地理环境教育部重点实验室,南京210023 [4]江苏省地理信息资源开发与利用协同创新中心,南京210023 [5]江苏省地理环境演化国家重点实验室培育建设点,南京210023

出　　处：《科学技术与工程》2020年第30期12268-12278,共11页Science Technology and Engineering

基　　金：自然资源部城市国土资源监测与仿真重点实验室开放基金(KF-2019-04-003);江苏省研究生科研与实践创新计划(SJCX19_0381);国家自然科学基金(41631175);国家重点研发计划(2017YFB0503500);江苏省海洋科技创新专项(HY2018-3);江苏高校优势学科建设工程资助项目(164320H116)。

摘　　要：利用航空影像已经成为获取DSM(digital surface model)等测绘数字产品及快速三维建模的重要手段之一。而作为其中关键技术之一的密集匹配则是其中的重点和难点。近几年来随着计算机硬件的快速发展,深度学习方法在图像处理等领域取得了巨大的成功。基于深度学习的影像密集匹配算法层出不穷,并且相较于传统方法取得了更好的效果,但提高匹配精度、提升匹配效率仍是未来研究的目标。通过对目前具有代表性的技术方法进行回顾,按照基于图像块的相似性度量学习和端对端生成视差图两类研究思路,叙述了深度学习在影像密集匹配中的研究进展与趋势,总结了现有方法的优点与不足,以期为影像匹配的研究提供具有参考价值的文献综述。The use of aerial images has become one of the prominent methods for obtaining DSM(digital surface model)and other digital surveying and mapping products and reconstruction 3D modeling.As one of the cardinal technologies,dense matching is the key and difficult point.With rapid development of computer hardware in recent years,deep learning methods have achieved great successes in image processing and other fields.The deep-learning-based image dense matching algorithms emerge enormously recently and have achieved better results than traditional methods.However,the matching accuracy and efficiency remained to be improved in the future.Therefore,the current representative technical methods were reviewed and the research progress and trends of deep learning in image-dense matching were summarized based on two types of research ideas:image block-based similarity measurement learning and end-to-end disparity map generation.The advantages and disadvantages of the existing methods were commented to provide a literature review with a reference value for the study of image matching.

关 键 词：密集匹配 计算机视觉 深度学习 视差估计 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]