分块鲁棒主成分分析的撞击坑图像检测识别 - 北京邮电大学学报.pdfVIP

2016年2月 北 京 邮 电 大 学 学 报 Feb.2016 第39卷 第1期 Journal of Beijing University of Posts and Telecommunications Vol.39 No.1 摇 摇 文章编号:1007鄄5321(2016)01鄄0063鄄05 DOI:10.13190/ j.jbupt.2016.01.011 分块鲁棒主成分分析的撞击坑图像检测识别 刘摇 安,摇 周东华,摇 陈茂银 (清华大学 自动化系,北京 100084) 摘要:针对遥感图像地形背景复杂的问题,提出分块鲁棒主成分分析的撞击坑候选区域自动提取方法. 基于图像 分块,采用交替方向乘子算法进行结构稀疏的低秩分解,低秩成分表示冗余相似的背景,稀疏成分代表包含潜在撞 击坑的显著区域. 针对显著的区域图采用数学形态运算分割获取候选的撞击坑图像,并通过对候选图像进行稀疏 表示的分类,识别出真实撞击坑. 基于火星和月球图像的实验结果表明,该方法能有效去除复杂地形和光照的干 扰,检测率达到91郾7%. 关摇 键摇 词:撞击坑检测;鲁棒主成分分析;视觉显著性;撞击坑候选区域 中图分类号:TP391郾41摇 摇 摇 摇 文献标志码:A A Robust Crater Detection and Recognition Method Based on Blocked Principal Components Analysis LU An,摇 ZHOU Dong鄄hua,摇 CHEN Mao鄄yin (Department Automation,Tsinghua University,Beijing 100084,China) Abstract:Crater is important for analyzing the relative dating of planetary and lunar surfaces. For the complex terrains in remote sensing images,a robust blocked principal components analysis (RPCA) ap鄄 proach wasproposed to automatically detect crater candidate regions. An alternating direction multipliers algorithm waspresentedfor RPCAbasedontheblockedplanetary images. Thebackgroundismodeledas a low鄄rankmatrix,andthesalientregionsmapisrepresentedby structuresparsepartsthatcontainpoten鄄 tial craters. The crater candidatesare obtained by mathematical morphological operationsfor the saliency regions map,they are precisely distinguished from falsely detected ones through a sparse representation classifier in feature space. Experimentsontheimagesfrom Marsand Moon demonstrate show thattheac鄄 curacy rate of crater recognition can reach up to 91郾7% by effectively eliminating the ef