基于NSCT、KFCM和多模型LS-SVM的红外小目标检测  

Small target detection in infrared image based on NSCT,KFCM and multi model LS-SVM

作　　者：吴一全[1,2] 尹丹艳[1] 吴诗婳[1] 

机构地区：[1]南京航空航天大学电子信息工程学院,南京210016 [2]南京大学计算机软件新技术国家重点实验室,南京210093

出　　处：《仪器仪表学报》2011年第8期1704-1709,共6页Chinese Journal of Scientific Instrument

基　　金：国家自然科学基金(60872065);航空科学基金(20105152026);计算机软件新技术国家重点实验室开放基金(KFKT2010B17)资助项目

摘　　要：为了进一步提高红外小目标的检测性能,针对图像序列中背景与小目标的特点,提出了一种基于非下采样Contourlet变换(nonsubsampled contourlet transform,NSCT)和核模糊C均值(kernel fuzzy C means,KFCM)聚类多模型最小二乘支持向量机(least squares support vector machine,LS-SVM)背景预测的检测方法。首先对红外小目标图像进行NSCT并去噪,提高图像的信噪比;然后通过基于核模糊C均值聚类的多模型LS-SVM预测去噪后红外图像中的背景,用去噪后的实际图像减去背景预测图像得到残差图像;接着提出基于递归最大类间绝对差的阈值选取算法分割残差图像;最后利用目标灰度的平稳性和运动轨迹的连续性进一步检测出真实的小目标。给出了实验结果与分析,并与现有的3种基于背景预测的小目标检测方法进行了比较。结果表明该方法具有更高的检测概率和信噪比增益。To further improve the detection performance of small infrared target, aiming at the characteristics of back- ground and small target in infrared image sequences, a detection method using nonsubsampled contourlet transform （NSCT） and background prediction based on the kernel fuzzy C means （KFCM） clustering and multi model least squares support vector machine （LS-SVM） is proposed. First, the infrared image is decomposed using NSCT and the noise is removed to increase signal-to-noise ratio. Then, multi model LS-SVM based on kernel fuzzy C means cluste- ring is adopted to predict the background of de-noised infrared image. The predicted background image is subtracted from the de-noised source image and a residual image is obtained. Next, a threshold selection method based on re- cursive maximum between-cluster absolute difference is presented to segment the residual image. Finally, the real small target is detected based on the stability of target gray and the consistency of target trajectory. Experimental re- sults and analyses are given, which are compared with those of three existing small target detection methods based on background predication. Comparison results show that the proposed method has higher detection probability and gain of signal-to-noise ratio （GSNR）.

关 键 词：红外小目标检测 非下采样CONTOURLET变换 核模糊C均值聚类 最小二乘支持向量机 递归最大类间绝对差 

分 类 号：TN911.73[电子电信—通信与信息系统]