低秩和稀疏分解的立体影像匹配错误点检测  

Image matching error point detection based on low-rank and sparse decomposition

作　　者：张正鹏[1] 张强 ZHANG Zhengpeng;ZHANG Qiang(School Geomatics,Liaoning Technical University,Fuxin.Liaoning 123000,China;Zhuhai Orbita Aerospace Science&Technology Co.,Ltd.,Zhuhai,Guangdong 519000,China)

机构地区：[1]辽宁工程技术大学测绘与地理科学学院,辽宁阜新123000 [2]珠海欧比特宇航科技股份有限公司,广东珠海519000

出　　处：《中国矿业大学学报》2020年第3期595-601,共7页Journal of China University of Mining & Technology

基　　金：国家自然科学基金项目(41801294);辽宁省自然科学基金项目(20180551209)。

摘　　要：提出一种基于低秩和稀疏分解的立体影像匹配错误点检测方法.以正确匹配点的运动结构相似性和错误匹配点的稀疏性为约束条件,考虑局部最近邻匹配点间的运动低秩特征,引入图拉普拉斯约束项来表达这种局部低秩性,在此基础上构建影像间匹配点的低秩和稀疏优化模型.采用自适应惩罚的线性化交替方向法推导并完成模型的低秩和稀疏分解.统计稀疏矩阵奇异值特征,以此为判断条件进行正确与错误匹配点的检测.实验选择具有高度纹理重复特征的立体像对,分别对比和分析不同转角下影像错误匹配点的检测精度.结果表明:在高纹理重复特征、高重叠度下,提出的方法较经典方法能更好的区分正确与错误匹配点,在正确率、召回率、精度和F值(F-measure)指标评价方面表现占优.A method of stereo image matching error point detection based on low rank and sparse decomposition was proposed.Based on the constraints of the similarity of motion structure and the sparsity of error matching points,and considering the low rank feature of motion between local nearest matching points,the constraint term of laplacian eigenmaps was introduced in order to express this local low rank property,on this basis,a low rank and sparsity optimization model of image matching points were constructed.Linearized alternating direction method with adaptive penalty was used to derive and complete the low rank and sparse decomposition of the model.The feature of singular value of sparse matrix was counted to complete the detection of correct and wrong matching points.The experiment choosed stereo image pairs with high texture repetition characteristics,and compared and analysed the detection accuracy of image error matching points under different rotation angles.The results show that under the condition of high texture repetition and high overlap,the proposed method could distinguish correct and wrong matching points better than the classical method,and is superior in accuracy,recall,accuracy and F-measure index evaluation.

关 键 词：低秩 稀疏分解 运动结构特征 影像匹配 

分 类 号：P237[天文地球—摄影测量与遥感]