优化稀疏表示的迷彩伪装目标高光谱识别方法研究  

A Hyperspectral Recognition Method for Camouflaged Targets Based on Background Dictionary Sparse Representation

作　　者：徐景余 包妮沙[1,2] 郎洁双 刘善军 毛亚纯[1,2] 贺黎明 XU Jing-yu;BAO Ni-sha;LANG Jie-shuang;LIU Shan-jun;MAO Ya-chun;HE Li-ming(School of Resources and Civil Engineering,Northeastern University,Shenyang 110819,China;Liaoning Advanced Technology Collaborative Innovation and Application Center for High-Resolution Earth Observation System,Shenyang 110819,China;Military-Civilian Integration Development Committee of the CPC Liaoning Provincial Committee,Shenyang 110032,China)

机构地区：[1]东北大学资源与土木工程学院,辽宁沈阳110819 [2]高分辨率对地观测系统辽宁先进技术协同创新与应用中心,辽宁沈阳110819 [3]中共辽宁省委军民融合发展委员会办公室,辽宁沈阳110032

出　　处：《光谱学与光谱分析》2024年第12期3534-3542,共9页Spectroscopy and Spectral Analysis

基　　金：国家自然科学基金项目(52074063);教育部联合基金项目(8091B032232)资助。

摘　　要：目前利用稀疏表示算法提取高光谱影像中伪装目标时,背景字典的选择受到伪装目标“异物同谱”的影响,导致无法准确检测伪装目标。为探索解决这一问题,以草地伪装网、荒漠伪装网为研究对象,分别采集伪装网可见光-近红外反射率光谱和机载高光谱影像,分析室外测得的伪装网、机载影像中背景像元与伪装目标像元的光谱特征。利用室外测得的伪装网与机载影像中的背景光谱具有差异性、相邻像元属于同一地物可能性大的特点,提出基于欧式距离和影像同质性特征约束的背景字典选择方法,进一步利用背景字典稀疏表示方法识别目标。结果表明:(1)对于草地伪装:在750~1000 nm波长范围内,影像中背景像元光谱的反射率高于伪装网光谱的反射率。对于荒漠伪装:在550~700 nm范围内,影像中背景像元光谱反射率高于伪装网光谱的反射率。(2)通过建立与目标像元光谱欧式距离最大,且与邻域像元同质性最高的空谱特征约束条件,选取了草地伪装影像413个背景像元和荒漠伪装影像507个背景像元作为背景字典。(3)基于改进的背景字典选择方法,利用稀疏表示算法对伪装目标进行识别,结果可以精确判别伪装目标的位置与数量。对草地伪装目标与荒漠伪装目标检测的接受者操作特征的曲线下面积(AUC)分别达到了0.96和0.98,表明该算法针对草地伪装目标、荒漠伪装目标均具有良好的检测性能。Referring to the current use of sparse representation algorithms to extract camouflaged targets from hyperspectral images,the selection of the background dictionary is affected by the“same spectrum of different objects”of the hidden targets,resulting in the inability to detect camouflaged targets accurately.In this paper,we take the grassland camouflage net and desert camouflage net as the research objects,collect the visible-near-infrared reflectance spectra of the camouflage net and airborne hyperspectral images respectively,and analyze the spectral characteristics of the background pixels and camouflage target pixels in the camouflage net and airborne images measured outdoors.Taking advantage of the fact that the spectra of the camouflage nets measured outdoors and the background in the airborne images are different and the possibility of neighboring image elements belonging to the same feature is high,the background dictionary selection method based on the constraints of Euclidean distances and image homogeneity features is proposed,and the sparse representation of the background dictionary is further utilized to identify the target.The results show that(1)in the wavelength range of 750~1000 nm for grass camouflage,the reflectance of the background pixel spectrum in the image is higher than that of the camouflage net spectrum.For desert camouflage:in the range of 550~700 nm,the reflectance of the background pixel spectrum in the image is higher than that of the camouflage net spectrum.(2)By establishing spatial and spectral feature constraints with the maximum spectral Euclidean distance to the target image element and the highest homogeneity with neighboring image elements,413 background image elements in the grass camouflage image and 507 background image elements in the desert camouflage image were selected as the background dictionary.(3)Based on the improved background dictionary selection method,the sparse representation algorithm is utilized to identify the camouflage targets,and the results can ac

关 键 词：机载高光谱 可见光-近红外光谱 伪装网 背景字典 稀疏表示 

分 类 号：TP79[自动化与计算机技术—检测技术与自动化装置]