基于孪生残差神经网络的遥感影像变化检测  

Remote sensing image change detection based on twin residual neural network

作　　者：倪良波 卢涵宇 卢天健 丁蕾锭 卢梅 NI Liang-bo;LU Han-yu;LU Tian-jian;DING Lei-ding;LU Mei(College of Big Data and Information Engineering,Guizhou University,Guiyang 550025,China;College of Resources and Safety Engineering,Central South University,Changsha 410083,China)

机构地区：[1]贵州大学大数据与信息工程学院,贵州贵阳550025 [2]中南大学资源与安全工程学院,湖南长沙410083

出　　处：《计算机工程与设计》2020年第12期3451-3457,共7页Computer Engineering and Design

基　　金：国家自然科学基金项目(41671355);国家重点研发计划基金项目(2016YFFE0117300);贵州省基金项目([2020]1Y155、[2020]4001、[2017]5788、2816、2025、[2016]2844、2801、5604、5803、[2015]3054)。

摘　　要：为减少遥感影像变化检测方法中“伪变化”的影响以及检测效果不理想等问题,提出一种基于孪生残差神经网络的变化检测方法。对多时相多光谱影像超像素进行分割与合并,对分割的子块提取特征,得到初级变化检测图,选择变化图中变化与未变化的区域为训练样本,通过孪生残差神经网络(SiameseResNet)进行二次分类获得相似度,经过OTSU阈值分割后获得到最后的变化检测结果。实验结果表明,超像素分割与二次分类的方法可以有效提高变化检测正确率,减少“伪变化”对变化检测的影响,具有较强鲁棒性。To reduce the influence of pseudo-change in traditional remote sensing change detection and to alleviate the effects of unideal change detection result,a change detection method based on Siamese Residual neural network was proposed.Multi-temporal remotely sensed image superpixels were segmented and merged,and the features of the segmented sub-blocks were extracted to obtain the primary change detection map.The changed and unchanged regions in the change map were selected as training samples,and the similarity was obtained through the secondary classification using SiameseResNet.The final change detection result was obtained by OTSU threshold segmentation.Experimental results show that the method of secondary classification and superpixel segmentation can effectively improve the correct rate of change detection,and effectively reduce the impact of pseudo-change on change detection,meanwhile,the results show that the method has strong robustness.

关 键 词：遥感影像 二次分类 变化检测 孪生残差神经网络 超像素分割 特征变化图 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]