基于图像序列稀疏表示的城市背景红外弱小目标核检测算法研究

项目名称： 基于图像序列稀疏表示的城市背景红外弱小目标核检测算法研究

项目编号： No.61307025

项目类型： 青年科学基金项目

立项/批准年度： 2014

项目学科： 无线电电子学、电信技术

项目作者： 朱斌

作者单位： 中国人民解放军电子工程学院

项目金额： 29万元

中文摘要： 红外搜索跟踪是城市防空预警的重要方式。城市环境复杂，建筑、树木、地形起伏等背景呈现明显的非线性、非平稳分布特性，而专门针对其环境特点进行的空中红外弱小目标检测算法的研究很少。该项目提出对红外图像序列进行稀疏表示，只用少量数据集中描述图像序列的重要特征，将能量集中于目标与城市环境下易产生虚警的各类非线性背景等感兴趣区域；然后，通过核方法将原始线性不可分的图像数据映射到高维特征空间，在特征空间寻找简单的线性分类方法来高效抑制非线性背景。主要研究内容为：(1)基于核方法的红外弱小目标检测算法；(2)揭示不同类型城市背景中各类图像信息及其数据冗余度对检测的影响机理与内在关联，探索红外图像序列的可压缩性；(3)设计图像数据稀疏准则，研究对城市背景红外图像序列的高效稀疏方法。该研究将推动核方法与图像压缩编码理论在红外搜索跟踪系统检测算法中的应用，也将有助于城市防空预警信息处理技术的发展。

中文关键词： 城市背景；红外图像序列；稀疏表示；弱小目标检测；核方法

英文摘要： Infrared search and track (IRST) is an important technology for urban air defense warning, but there is almost no research focus on the aerial dim target detection in city background yet. The complex environment of city make the image background such as buildings, trees, landform undulation, and etc present an obvious nonlinear and non-stable distribution. An infrared image sequence sparse representation method is proposed in this project. With this method, the important character of image sequence is described only by much less information, the energy of image is concentrated on target and those interested non-linear city background areas which often cause false alarm. After that, map the unclassifiable original image data to a higher eigen space, in order to find a simple linear algorithm to suppress the nonlinear city background efficiently. This project will mainly focus on: (1) infrared dim target detection algorithm based on kernel methods. (2) Study on the influence and internal relations between target detection and different kinds of information and its redundancy in different kinds of city background; and the compressibility of infrared image sequence. (3) Design a sparsification rule to classify and distinguish image data, research an efficient sparse algorithm to infrared image sequence in city envir

英文关键词： city background；infrared image sequence；sparse representation；dim target detection；kernel methods