自适应实时红外小目标图像处理方法研究

When adopted to analyze the images of infrared small targets with low signal-to-noise ratio against complex backgrounds, the traditional localization methods restricted to the Heisenberg Uncertainty Principle expose the problem of mode mixing and bad self-adaptive property, and fail to meet the requirement for the real time. Consequently, the project utilizes the self-adaptive decomposition property of empirical mode decomposition (EMD), aiming to study a novel self-adaptive and real-time processing method of infrared small target images. Firstly, correctly understand the problem of mode mixing of infrared small target images against complex backgrounds, and then propose an effective de-mixing method based on the multi-frame integration. Secondly, explore a realization method of bidimensional EMD, then construct a self-adaptive image preprocessing algorithm and a self-adaptive segmentation method to enhance the capability of the candidate target extraction. Thirdly, consider the spatial and temporal features of the targets, and study a true or false target discrimination method based on the temporal-spatial fusion filtering, aiming to lessen the computation cost and meet the requirement for the real time. Meanwhile, develop a simulation platform for processing algorithms of infrared small target images, aiming to improve and perfect the theoretical research based on the simulated results. The study provides the design and development of the self-adaptive and real-time processing system of infrared small target images with key theoretical bases and application schemes. Therefore, the project enlarges the practical application of the processing technology of infrared small target images in military and civil fields.

传统局域化分析方法因受Heisenberg测不准原理的制约，在分析复杂背景下低信噪比红外小目标图像时,易产生模态混叠问题，自适应性差，且难以满足系统实时性的要求。针对上述挑战，本项目基于经验模态分解（EMD）自适应局域化分析特性，探索复杂背景下自适应实时红外小目标图像处理新方法。首先，正确认识复杂背景下红外小目标图像的模态混叠及机理，并设计基于多帧集成的去混叠算法，提升算法分析性能；其次，探索符合红外图像特性的二维EMD方法，并依此构建图像自适应预处理、自适应分割等技术，高效提取候选目标；最后，结合目标的时空特性，研究基于时空域融合滤波的真假目标甄别方法，降低算法运算量，满足系统实时性需求。与此同时，开发红外小目标图像处理算法仿真平台，并通过仿真改进和完善理论研究。本项目为自适应实时红外小目标图像处理系统的设计和开发提供关键理论基础和应用方案，拓展该技术在军用和民用领域的应用范围。

小目标图像处理技术已在军事领域（如精确制导、预警探测、战地指挥和侦察、敌我识别等）和民用领域（如卫星大气红外云图分析、医疗图像病理分析、地质分析、入侵检测、森林火灾探测等）得到持续的普遍关注。但是，小目标图像中的目标小、强度弱，没有先验的大小、形状、纹理及运动等信息，且目标、背景和噪声混叠在一起，难以直接检测。传统局域化分析方法受Heisenberg测不准原理制约，易产生模态混叠问题，且自适应性差，导致小目标检测的虚警率和漏检率偏高。本项目针对上述挑战，发展自适应局域化红外小目标图像分析及检测方法。项目总体取得了良好的研究成果：1）由于小目标的出现只引起图像全局纹理的较小变化，但会引起局部区域纹理的较大变化，我们提出了一种加权局部图像熵抑制红外图像背景和噪声、增强目标，最终大幅度地提高图像信噪比；2）为有效度量红外图像中目标区域与邻域背景之间的灰度差异，设计了一种基于加权局部差异度量的小目标检测方法，该方法可以简单、有效地检测不同复杂背景下的红外小目标；3）针对传统的基于局部变化的小目标检测方法可能导致较高的漏检率和虚警率，构建了一种基于熵的窗口选择方法，自适应选择局部区域尺寸，有效避免目标漏检问题；4）依据模糊集理论，构思了一种模糊距离概念，通过度量目标内部、背景内部、目标与背景之间的模糊距离，将因小目标的出现导致的局部纹理变化转化为模糊距离的度量，进一步提高红外小目标的检测率及降低虚警率。本项目的实施为自适应红外小目标图像处理系统的设计和开发提供理论基础和应用方案，拓展该技术在军用和民用领域的应用范围。项目执行期间共发表署课题号论文10余篇SCI论文，包括IEEE Trans. Cybernetics, IEEE Trans. Geoscience & Remote Sensing, Pattern Recognition, IEEE Trans. Aerospace & Electronic Systems, IEEE Trans. Biomedical Engineering等，6个授权专利，培养博士生3名。