采用监督特征学习的红外小目标检测  

Small infrared target detection via supervised feature learning

作　　者：许庆晗[1] 金立左[1] 费树岷[1] 

机构地区：[1]东南大学自动化学院,南京210096

出　　处：《东南大学学报（自然科学版）》2011年第5期1008-1012,共5页Journal of Southeast University：Natural Science Edition

基　　金：航空科学基金资助项目(20080169003)

摘　　要：为了改善大尺寸图像下红外小目标检测的检测率与速度,提出一种采用监督特征学习的检测算法.通过分析小目标邻域图像的分布特点,定义一种基于灰度分布的统计特征,用以描述目标与非目标的邻域的灰度分布差异.以局部灰度极大值区域为训练样本,通过有监督学习提取对目标区分能力最强的特征.随后,在特征空间设计级联结构的多分类器,采用逻辑斯蒂回归和相关向量机分类器,通过"目标-非目标"分类,实现对目标的检测.实验结果表明在相同虚警率下,检测率较局部滤波法有一定提升,且检测速度大幅提高,满足了大尺寸图像下的实时性要求.A supervised feature learning method is proposed for improving the detection probability and detection speed of small infrared target detection.Through analyzing the traits of small targets＇ neighborhood image,a statistical feature based on gray intensity distribution is defined for describing the difference between the targets and nontargets＇ neighborhood.Intensity extrema on global images are considered as training samples,and then a feature with the highest discriminability is extracted by supervised learning.Subsequently,a multi-stage classifier is designed in the feature space,which adopts logistic regression and relevance vector machine algorithms to detect targets via ＂target-nontarget＂ classification.Experimental results indicate that for large scale images and with the same false alarm rate,the proposed method is of higher probability of detection and much faster detection speed than local filtering methods.

关 键 词：小目标检测 灰度分布特征 相关向量机 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]