【作者】 尹丹艳；

【导师】 吴一全；

【作者基本信息】 南京航空航天大学 ， 信号与信息处理， 2010， 硕士

【摘要】 红外小目标检测是预警、搜索跟踪和红外自动寻的制导等领域的关键问题。如何提高小目标检测的可靠性和准确性,在红外探测成像方面具有重要的研究意义。本文在综述国内外小目标检测技术发展现状的基础上,研究了基于背景预测的红外小目标检测方法,主要工作如下:1、研究了基于双树复数小波变换(DT-CWT)、支持向量回归(SVR)的红外小目标检测方法和基于模糊最小二乘支持向量机(FLS-SVM)的红外小目标检测方法。前者采用了DT-CWT对红外图像进行去噪,并用SVR预测背景;后者利用FLS-SVM实现了红外图像背景的预测。实验结果表明,这两种方法都能较精确的检测出红外小目标,检测概率较高,检测效果均优于现有的红外小目标检测方法。2、提出了基于非下采样Contourlet变换(NSCT)和核模糊聚类(KFCM)多模型LS-SVM的红外小目标检测方法。该方法首先对红外图像进行NSCT去噪,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,最后分割残差图像并利用小目标的运动特性检测出真实小目标。实验结果显示该方法具有更高的检测概率和信噪比增益。3、给出了基于混沌粒子群(CPSO)和最小一乘空时预测的红外小目标检测方法。首先建立基于最小一乘准则的空时背景预测模型,然后根据最小一乘估计的性质,提出应用CPSO解决最小一乘估计中的极值选取问题,最后用该模型预测红外图像中的背景并从残差图像中分割检测出小目标。结果表明该方法优于基于最小二乘背景预测的红外小目标检测方法。4、实现了基于灰色预测的红外小目标检测。应用灰色系统理论中的GM(1.1)模型对红外小目标图像中的背景进行时域预测,最终从残差图像中分割出红外小目标。实验结果表明,该方法可以实现较远距离小目标的检测。5、研究了红外小目标残差图像的阈值分割算法。重点研究了基于模糊Tsallis-Havrda-Charvat熵的阈值选取算法、基于递归最大类间绝对差的阈值选取算法和基于CPSO的二维直方图斜分模糊最大熵阈值选取算法。实验结果表明这些阈值选取算法选取的阈值能从红外小目标残差图像中准确地分割出小目标。更多还原

【Abstract】 The detection of small infrared target is the key problem in areas such as early warning, search and tracking, and infrared automatic homing guidance. How to improve the reliability and accuracy of small target detection has much research significance in infrared detection imaging. On the basis of introducing domestic and foreign development of small target detection, this paper studies the methods of small infrared target detection based on background prediction. The main tasks are as follows:Firstly, two detection methods of small infrared target are studied. One is based on dual tree complex wavelet transform (DT-CWT) and support vector regression (SVR). And the other is based on fuzzy least squares support vector machines (FLS-SVM). The former method suppresses the noise in infrared image by DT-CWT and predicts the background by SVR. And the latter one adopts FLS-SVM to predict the background of infrared image. The experimental results show that both methods can detect the small infrared target accurately and have high detection probability. Their detection results are better than the existing methods of small infrared target detection.Secondly, a detection method of small target in infrared image using nonsubsampled contourlet transform (NSCT), kernel fuzzy C means (KFCM) clustering and multi model LS-SVM is proposed. First, the infrared image is de-noised by NSCT. Then, the multi model LS-SVM based on KFCM clustering is adopted to predict the background of the de-noised infrared image. Finally, the real small taget is detected by segmenting the residual image and using the motion characteristics of small target. The results show that this method has higher detection probability and gain of signal-to-noise ratio (GSNR).Thirdly, a detection method of small infrared target is proposed, which is based on chaotic particle swarm optimization (CPSO) and spartial-temporal background prediction by least absolute deviation. First, a model of spatial-temporal background prediction is built. According to properties of least absolute deviation, the extreme value in least absolute deviation is selected by CPSO. The background in infrared image is predicted by this model and the small target is detected by segmenting the residual image. The results show this method is superior to the method of small infrared target detection based on background predication by least squares.Fourthly, a detection method of small infrared target based on gray prediction is realized. The GM(1.1) model of gray system theorey is adopted to predict the infrared image background in time domain. The small target can be detected by segmenting the residual image. The experimental results show that this method can achieve long-range small target detection.Fifthly, some threshold segmentation algorithms for the small infrared target of residual image are studied, such as the threshold selection method based on fuzzy Tsallis-Havrda-Charvat’s entropy, the threshold selection method based on recursive maximum between-cluster absolute difference and a two-dimensional histogram oblique segmentation method based on CPSO and fuzzy maximum entropy. The results show that the threshold selected by these methods can accurately segment the small target from infrared residual image.更多还原