【作者】 刘瑞明；

【导师】 刘尔琦；

【作者基本信息】 上海交通大学 ， 模式识别与智能系统， 2008， 博士

【摘要】 红外目标图像处理是早期预警、安全监控、目标搜索跟踪、导弹制导等军事和民用领域的一项非常关键的技术,许多科研机构的大批科研工作者致力于该项技术的研究。现代战争是高技术条件下的激烈对抗,这就要求武器系统具有远距离检测和跟踪目标的能力,以赢得更多的作出反应的时间,从而在战场中取得主动。当目标距离红外探测器较远时,红外目标的尺寸很小、对比度很低,没有明显的纹理、结构等特征,使目标的检测和跟踪相当困难。此外,由于通常情况下,目标所处的背景极其复杂,目标被大量出现的杂波和噪声所污染,使红外目标的处理更加困难。目前国内外学者已经提出了一些红外目标的检测和跟踪算法。传统的红外目标检测方法主要是基于高通或低通滤波和统计学滤波,后来又出现了基于二维最小二乘的检测方法,值得注意的是,随着模式识别技术的发展,基于该技术的检测方法正逐渐成为一个新的研究热点。对于跟踪问题,专门针对于红外目标的跟踪算法还没有出现,目前所用的算法都是借用彩色图像的目标跟踪算法,比如均值漂移算法(Mean shift)和粒子滤波算法(Particle filter),尽管这些算法在跟踪彩色图像时跟踪精度较高,但由于红外目标的特征较弱,实践证明,直接借用这些算法,跟踪效果并不理想。本文致力于探索新的红外小目标检测算法,以提高目标检测概率,系统地提出了基于子空间(包括线性子空间和非线性子空间)的红外点目标检测算法。此外,如何提高红外目标的跟踪精度也是本文的研究重点。一幅包含有目标的数字图像,可以被视为包含目标数据和背景数据的数据集,目标检测任务就是将目标数据从数据集中挑选出来。因此,目标检测问题可以被转化为一个两类模式的识别问题,任何一种模式识别的分类算法都可以用来完成目标检测任务。本文的着眼点是用子空间方法来完成红外目标的检测,研究子空间方法检测红外目标的可能性,并验证其性能,就此我们利用线性子空间理论提出了Eigentargets和Fishertarget的概念,并与另一种子空间方法FKT一起用于红外点目标的检测。线性子空间方法只能提取目标的二阶统计特性,为了能挖掘目标的高阶统计特性,线性子空间检测方法被推广到高维特征空间,开发出基于核方法的目标检测算法。因为原始红外目标的特征很弱(或很少),为了能使跟踪算法的精度更高,我们提出多特征图像的概念,首先提取图像的Gabor特征和熵特征,然后将这些特征生成的图像和灰度图像合成多特征图像,均值漂移算法在多特征图像上的跟踪精度远高于在原灰度图像上的跟踪精度。具体来说,本文有以下几个主要创新点:1.对于有监督的模式识别方法,训练样本的选择对识别率的影响非常大,只有合理的训练样本才能真正反映识别方法的性能。高斯灰度模型(Gaussian intensity model,GIM)是一种常用的描述红外点目标的数学工具,也被用来产生红外点目标的训练样本。但GIM容易产生样本外点(Outlier),它们会大大降低识别算法的识别率。为了提高红外点目标的检测性能,本文对高斯灰度模型进行改进,改进后的GIM产生的模拟红外点目标,更接近真实目标,而且避免了样本外点的产生,为开发基于模式识别的红外点目标检测方法奠定了基础。2.线性子空间方法已经在人脸识别领域得到了成功应用。但还没有人系统地提出用线性子空间方法来检测红外点目标。本文给出特征目标(Eigentargets,Ets)和费舍尔目标(Fishertargets,Ft)的概念,通过投影所得坐标与目标训练样本的投影坐标之间的距离(范数)或重构误差来完成目标的检测。开发出一种目标检测函数来生成检测图像,使检测算法更方便于性能上的比较研究。为了进一步提高目标检测性能,将所提出的线性子空间检测算法推广到高维特征空间,形成基于核方法(非线性子空间)的红外点目标检测算法。线性子空间方法只能挖掘图像的二阶统计特性,在检测目标时,二阶统计特性不足以描述目标图像和背景图像的特征,而核方法能挖掘图像的高阶统计特性。因此,基于核方法的红外目标检测方法,能够获得更好的检测性能。3.红外点目标图像是灰度图像,目标的特征极不明显,几乎没有纹理等特征,边缘强度也很弱,均值漂移算法和粒子滤波等跟踪方法跟踪红外点目标时,容易丢失目标,或跟踪精度不高。我们在所提出的子空间方法的基础上,结合Kalman预测技术,开发出一种红外点目标跟踪方法,获得了较高的跟踪精度。此方法也可以认为是一种检测前跟踪(TBD)的多帧检测算法。4.对于红外目标的跟踪,目前还没有开发出专门针对红外面目标的跟踪算法,而由于红外面目标的统计特征较弱,用均值漂移算法来进行跟踪,效果不能令人满意。本文提出一种基于多特征图像的均值漂移跟踪算法,将Gabor特征图像和熵特征图像与原始的红外灰度图像一起合成多特征图像。在多特征图像中,目标和背景具有更大的可区分性。均值漂移算法在多特征图像上对红外面目标进行跟踪,算法对目标更多的特征进行模式匹配,提高了跟踪的精度。本方法具有开创性,提供了一种新的思路,使均值漂移算法具有了处理目标其它特征的能力,弥补了此算法的一大缺陷。更多还原

【Abstract】 Infrared image processing is a key technique in the application field of early warning, safety monitoring and control, automatic search and tracking, missile guidance etc. Many researchers commit themselves to the study on this technique. The modern battles are intense confrontation with advanced technologies. For gaining dominance on the field of battle, the arm systems have to detect and track targets from far distance in order to save more time for dealing with the complexion of the war. Under this condition, infrared targets have small size and poor contrast without texture and configuration features. Therefore the detection and tracking of these targets are very difficult. Moreover, the background of infrared targets is complex with much clutter and noise, which makes it more difficult to dectect and track targets.Some methods for decting and tracking infrared targets have been proposed so far. Classical methods are based on high-pass or low-pass filters and order-statistical filters. The tow-dimensional least mean square filters is another method based on filters. With the development of pattern recognition, detecting targets with this theory is becoming a new research field. For tracking targets, the tracking methods designed specially for infrared targets are not proposed. So we have to use the methods designed for color images, such as mean shift and particle filter. Although they have good tracking precision for color images, their tracking effects are not satisfying for infrared images whose features are not distinct.This dissertation focuses on finding novel detection and tracking algorithms for infrared targets. One digital image with targets can be treated as a datum set consisting of target data and background data. Detecting targets is to find target data from the datum set. So the target detection problem is transformed into a tow-pattern recognition problem and any pattern recognition method can be used to detect targets. In this dissertation, we mainly use sub-space methods to detect targets and propose the concepts, Eigentargets and Fishertarget. Moreover, another sub-space method, FKT, also is researched on its detection performance. The linear sub-space methods are appropriate for data that are generated by Gaussian distribution, or data that are best described by second-order correlations without taking higher-order statistics into account. For capturing higher-order dependencies in the data, the nonlinear sub-space detection methods are proposed. The features of original infrared images are poor and faint. For improving the tracking precision, we track infrared targets in multiple feature images. They are synthesized by Gabor feature images, entropy feature images and original infrared images. Because the distinction between targets and background is improved, the mean shift can gain better tracking performance.Specifically, the main contributions of this dissertation are as follows:1. For supervised learning methods, training samples are very important. The appropriate training samples can correctly reflect the performance of recognition methods. The Gaussian intensity model (GIM) is a commonly used model to represent infrared point targets and applied to create training samples, simulated images, of infrared point targets. However, it may produce sample outliers which will reduce the recognition rate. We modify the GIM to avoid producing outliers. This is a foundation of exploring the detection methods, based on pattern recognition theory, for infrared point targets.2. Linear sub-space methods have been successfully applied to the field of face recognition. However, no one systematicly proposes the detection methods for infrared point targets which are based on linear sub-space methods. The concepts of Eigentargets and Fishertarget are given in this dissertation. The difference between projecting coefficients and the projecting coefficients of training samples of targets can indicate the resembling degree of the image and targets. So we can find targets from images. Moreover, we give a detection function to produce detection images which makes the performance comparison of detection methods be more convenient. To further improve detection performance, the linear sub-space methods are expanded into kernel methods (nonlinear sub-space methods) to detect infrared point targets. The linear sub-space methods can only capture the second-order statistics which is not sufficient for representing target images and background images. However, kernel menthods can capture the higher-order statistics. So the detection algorithms based on kernel methods have better detection performance.3. The images of infrared point targets are intensity images. Their features are not distinct. There is no texture features and the edge is dim. When the mean shift and particle filter algorithms track these targets, they often lose targets or the tracking precision is not high. We combine the proposed sub-space detection algorithms with Kalman prediction to track infrared point targets. This method has a high tracking precision. It can also be considered as a TBD method.4. At present, the tracking methods specially designed for infrared surface targets are not found. The mean shift can not gain satisfying performance when tracking them because of the poor feature. For resolving this problem, we propose a novel method for tracking infrared surface targets. The Gabor feature and entropy feature first are extracted from original infrared images. Then the feature images are used to synthesize pseudo-color image which is named as multiple feature image (MFI) by us. Thus the distinction between targets and background is more appreciable in MFI. The targets are tracked in MFI by mean shift which can complete the mode search with more features. So the tracking precision becomes higher. This method provides a novel idea which makes mean shift have a capacity of capturing more features besides the statistical feature of pixel intensity. It indirectly removes the deficiency of mean shift.更多还原