摘要
在目前现存的数字地表模型生成方法中,大多采用半全局立体匹配(semi-global matching,SGM)及其派生算法来进行遥感影像的密集匹配。传统SGM的匹配结果总体上较好,但仍存在一些不足,例如在阴影区域、低纹理、重复纹理以及局部光强不一致等区域存在较多的视差空洞,且在视差不连续的区域易存在较大的匹配误差等。近年来,基于深度学习的密集匹配方法在多个数据集上取得了较好的成绩。该文将孪生神经网络计算匹配代价引入资源三号密集匹配生成数字地表模型流程中,实验了深度学习方法在国产资源三号02星密集匹配方面的性能;对模型泛化能力进行了针对训练,和经典方法进行了比较,并与商业软件进行了精度对比。实验表明,与传统的影像密集匹配生成数字地表模型的方法相比,基于深度学习的匹配效果更优。
In the current existing digital surface model generation methods,semi-global stereo matching and its derived algorithms are mostly used for dense matching of remote sensing images.The matching results of traditional semi-global matching(SGM)are generally good,but there are still some shortcomings.For example,there are more parallax holes in areas such as shadows,low textures,repeated textures,and local light intensity inconsistencies,and large matching errors are likely to occur in areas where the parallax is discontinuous.In recent years,the dense matching method based on deep learning has achieved good results on multiple data sets.In this paper,the siamese neural network is introduced to calculate the matching cost in the stereo matching of ZY-3 satellite to generate digital surface model.It discusses the performance of deep learning in the dense matching of domestic ZY-302 satellite,and compares it with classic methods,targeted training on the generalization ability of the model.Compared with the classic method and the accuracy of commercial software,the experiment shows that the current deep learning matching effect is better.
作者
韩昊
李参海
丘晓枫
HAN Hao;LI Canhai;QIU Xiaofeng(School of Geomatics,Liaoning Technical University,Fuxin,Liaoning 123000,China;Land Satellite Remote Sensing Application Center,Ministry of Natural Resources of the People’s Republic of China,Beijing 100048,China)
出处
《遥感信息》
CSCD
北大核心
2022年第3期101-108,共8页
Remote Sensing Information
基金
国家重点研发计划战略性国际科技创新合作重点专项(2016YFE0205300)。
