导　　师： 秦前清;王文伟

学科专业： 081002

授予学位： 博士

作　　者： ;

机构地区： 武汉大学

摘　　要： 特定视场中红外成像目标检测技术的研究是利用红外光学成像传感器,从低空、深空、海天及陆地等客观场景中获取目标／背景信息,使用信号处理手段和基于视觉机制的图像识别理论与人工智能技术自动地分析理解场景信息,检测感兴趣的红外目标/(如侦察机、导弹、车辆、船舶和遇难人员等/),获取目标各种定性、定量特征,在线识别和判读目标,进而确定目标类型和锁定指定目标,实现自动化的智能监控。红外目标检测一直是计算机视觉领域中备受关注的前沿方向,复杂场景下的红外目标检测因其现有的低检测率和极高的实际应用价值,吸引着众多学者进行深入研究。本文依据调研情况,针对多种复杂红外场景,从三方面展开深入研究：一是远距离复杂天空背景下的红外弱小目标检测技术研究,二是复杂户外场景下单源红外目标检测技术研究,三是复杂户外场景下红外与可见光协同目标检测技术研究,取得的主要研究成果如下： 1、针对远距离复杂天空背景,实现有效抑制背景突显目标的方法。研究一种基于局部分析机制的红外图像增强方法,分析分数阶积分理论,研究分数阶积分的频域特性,结合红外图像特征,尝试性地将分数阶积分理论应用于红外弱小目标增强,提出一种基于分数阶积分算子的弱小目标增强算法。实验结果表明该方法能有效提高图像信噪比与对比度。 2、针对远距离复杂天空背景,以背景信息为出发点,提出一种基于人工免疫网络/(aiNet/)进行背景抑制、基于行列k均值/(k-means/)聚类实现阈值分割的单帧红外弱小目标检测方法。首先采用aiNet结合Robinson警戒环技术,融入自组织特征映射/(SOM/)石扑思想,设计一系列抗体进化策略,建立自适应局部空间背景模型一模糊拓扑记忆抗体库,并以此分析各像素点的背景模糊隶属度来抑制背景杂波;接着提出基于行列k-means聚类的阈值分割算法来检测真实目标。实验结果表明,该方法的F1指标高达99/%,其能随背景的局部变化来自适应建立空间背景模型,从而自适应抑制背景杂波突显目标,能有效提高信噪比检测弱小目标。 3、针对远距离复杂天空背景,以弱小目标信息为出发点,提出一种基于模糊自适应共振理论/(Fuzzy-ART/)进行背景抑制、基于行列模糊自适应聚类实现阈值分割的单帧红外弱小目标检测方法。首先依据红外成像原理,仿真生成红外弱小目标训练样本；然后采用Fuzzy-ART/(?/)申经网络建立目标模型,并以此分析各像素点的目标模糊隶属度来抑制背景杂波;最后采用基于行列模糊自适应聚类的自适应阈值分割算法来检测真实目标。实验结果表明,该方法能有效地抑制背景杂波突显目标,并能有效提高信噪比检测弱小目标。 4、针对民用复杂户外场景,搭建红外运动目标检测框架,实现高效的目标自动检测方法。基于统计信息分类思想,提出一种基于时空协同机制的红外目标检测框架：“背景提取-背景抑制-背景建模-目标定位-目标检测”。首先分析红外成像的时空相关性,采用多层次时空中值滤波器提取背景帧,接着采用主成份分析技术分解时域关联信息抑制背景杂波；然后采用空间关联模糊自适应共振神经网络建立时空背景模型定位目标；以此进行局部分类检测红外目标,最后采用基于纹理的二值限制型主动活动轮廓模型来提取准确的目标轮廓。实验结果显示,该方法的Fl指标高达96.3/%。该方法能有效抑制背景定位目标,有效检测出复杂场景下的红外目标。 5、针对民用复杂户外场景,协同红外与可见光数据源,建立可靠的红外与可见光协同目标检测方法。针对复杂场景下现有单一类型成像监控警戒系统所存在的适用范围窄、杂波干扰强和目标检测率低等问题,提出一种基于多值免疫网络模型的红外与可见光协同目标检测方法。首先采用模糊自适应共振神经网络建立红外与可见光各自的背景模型；接着依据多值免疫网络模型,将红外背景模型视为B细胞,可见光背景模型视为T细胞,设计一系列免疫应答策略来协同建立B细胞与T细胞的交互模型,并以此分析各像素点的背景模糊隶属度来检测目标。实验结果表明,该方法的F1指标高达96.4/%,能有效协同互补红外与可见光信息,检测出复杂场景下的目标。 The research of target detection from infrared imagery in specific scene can be illustrated as follow:The thermal imaging sensor is utilized to capture the information of target or background from actual scenes such as low sky, deep sky, sea-sky as well as land background; Then the signal and image processing theorys are combined with the artificial intelligence technology to analyze the scene information automatically and detect infrared targets of interest /(eg. spy plane, missile, vehicle, ship, wrecked people and so on/), whose some qualitative and quantitative features can be obtained; With these information, targets can be recognized onlinely and specified targets can also be locked, so as to realize the automatic intelligent surveillance. Nowadays infrared target detecion is a very active research area in computer vision. As the infrared target detection in complex scenes presents a low detection index, while it has high application value extremely, more and more researchers pay attention in this topic. Acoording to the investigation, in views of various complex infrared scenes, this thesis researchs three aspects of infared target detection:dim target detection in heavy clutter based on far-distance IR imagery, target detection in complex outdoors scene based on infared survlliance, and target detection in complex outdoors scene based on thermal-visible survlliance. The main contributions of this thesis can be concluded as follows. 1、Aim at the complex far-distance sky background, this thesis presents a novel enhancement approach for infrared dim target based on local analysis mechanism, which is capable of suppressing background clutters and highlighting targets. It studies the fractional integral theory, and analyzes its character in the frequency domain. Then according to features of infrared image, the fractional integral is utilized to enhance the infared dim target. Experiment results show that the proposed approach is capable of eliminating the background clutters and enhancing the signal-to-noise ratio effectively. 2、Aim at the complex far-distance sky background, in view of the background information, a novel infrared dim target detection approach is presented, which is based on background suppression by artificial immune network /(aiNet/) and threshold segmentation by k-means cluster of rows and columns. Firstly, the aiNet is combined with Robinson guard to build the adaptive local spatial background models as fuzzy topological memory antibody bank. In the process of antibody bank modeling, a series of antibody evolution strategies are designed based on self-organizing map /(SOM/). With these models, background clutters are suppressed according to the degree of fuzzy match between pixels and models. Then the proposed adaptive segmentation algorithm based on k-means cluster of rows and columns is utilized to detect the true targets. Experimental results show that the F1measurement of the proposed approach is up to99/%. The proposed approach is able to build the spatial background models adaptively according to the local change of image, and suppress the background clutters and highlight targets effectively. It is capable of improving the signal-to-noise ratio of images and detecting targets effectively. 3、Aim at the complex far-distance sky background, in view of the dim target information, a novel infrared dim target detection approach is presented, which is based on background suppression by Fuzzy adaptive resonance theory /(Fuzzy-ART/) and threshold segmentation by adaptive segmentation algorithm based on fuzzy cluster of rows and columns. Firstly, the infrared dim target training set is simulated according to the principle of thermal imagery. Then a Fuzzy-ART neural network is utilized to build the target models. With these models, the background clutters are suppressed according to the degree of fuzzy match between pixels and models. Lastly, the adaptive segmentation algorithm based on fuzzy cluster of rows and columns is utilized to detect the true targets. Experimental results show that the proposed approach is able to suppress background clutters and highlight targets effectively. It is capable of improving the signal-to-noise ratio of images and detecting targets effectively. 4、Aim at the complex outdoors surveillance scenes, a detection framework for infrared moving target is builded. In view of statistical classification, this thesis presents a multi-stage classification approach to detect the target based on a spatial-temporal detection framework:background extraction, background suppression, background model, target location and target detection. At first, a multi-level spatial-temporal median filter is utilized to extract the background frame, with which the background clutters are suppressed by using the principal component analysis technique. A spatially related Fuzzy ART neural network is then applied to identify the local regions-of-interest /(ROI/). Within each region, another Fuzzy ART neural network is utilized to detect the target. Lastly, a binary constrained texture-based active contour model is applied to extract each continuous silhouette. Experimental results demonstrate that the proposed approach is capable of detecting infrared moving targets and extracting the silhouettes effectively for F1measurement up to96.3/%. 5、Aim at the complex outdoors surveillance scenes, the thermal and visible imagery are combined to detect the target. In order to solve the narrow applicable range, the heavy clutters as well as the low detection ratio problems more effectively, this thesis presents a novel target detection approach in thermal-visible surveillance based on multiple-valued immune network. Firstly, two Fuzzy ART neural networks are utilized to build the background models of thermal and visible components. Then according to the multiple-valued immune network model, a series of immune response strategies are designed to cooperate B cell with T cell to build the interactive model, which takes the infrared background model as B cell, the visible background model as T cell. With the interactive model, the targets are detected according to the degree of fuzzy match between pixels and models. Experimental results show that the Fl measurement of the proposed approach is up to96.4/%. It is able to complement information between thermal and visible components effectively. The proposed approach is capable of detecting targets in complex scenes effectively.

关 键 词： 红外成像 目标检测 弱小目标 背景抑制 时空框架 人工免疫机制 协同检测

分 类 号： [TP391.41]

领　　域： [自动化与计算机技术] [自动化与计算机技术]