结合背景特性与目标正演的稀疏动态反演空间红外弱小目标检测方法

Spatial infrared imaging targets show a characteristic of blur/flicker, low contrast and signal-to-noise ratio/clutter (SNR/SCR), no concrete shape or texture feature. The existing target detecting pattern can’t meet the requirement of detection accuracy and system performance. This project presents a new theory and method of detecting infrared dim and small target, which combines forward modeling with sparse dynamic inversion, as well as involving both background and target features. Major creative research works include: (1) Study on the spectral inversion based high resolution reconstruction method for infrared images. This research aims at enhancing the detailed information and feature information of dim and small targets. (2) Study on the forward modeling problem of infrared images, including background and target feature research. We present imaging simulation and forward modeling research with different background and various target features, which provides prior constraints for target detection in complex real image sequences. (3) Study on the theory and method of infrared dim and small target detection based on sparse dynamic inversion. We present the process of target detection and recognition through low rank and sparse decomposition in single image, and combine the procedure of dynamic inversion under multi-frame constraints. This process is expected to reach the ability of successfully detecting long-distance imaging targets with the minimal size of 2×2 pixels, and the target contrast is not higher than 2dB. When the false-alarm rate is lower than 1e-3, this method can reach a detection probability of 90%, the processing speed is not lower than 30fps. This project establishes the entire theoretical framework of target detection based on forward modeling and image inversion, forming a novel technique for infrared dim and small target detection.

空间红外成像目标表现为模糊/闪烁、对比度低、几何纹理特征微弱，现有检测模式难于满足探测精度和性能要求。本项目提出背景与目标特性结合，正演模拟与稀疏动态反演结合的红外弱小目标检测理论及方法。主要创新研究包括：(1) 基于“谱”反演的红外图像高分辨细节重建方法，突出红外图像细节及弱小目标特征；(2) 背景与目标特性研究相结合的红外弱小目标正演模拟方法，包括不同背景/目标特性的成像仿真和正演模拟研究，为实际复杂场景的目标检测提供先验约束信息；(3) 基于稀疏动态反演的红外弱小目标检测理论及方法，通过对实际单帧图像的低秩与稀疏分解及多帧约束下的动态反演，进行目标检测与识别，实现最低2×2像素、对比度2dB的远距离红外弱小目标检测，达到1e-3虚警率下，检测率>90%，处理能力>30fps的指标要求。通过本项目研究，探索形成一种新的红外弱小目标检测理论及技术途径，解决目前空间目标探测所面临的难题。

本项目围绕远距离红外目标探测中的低对比度、低信噪比及信杂比、成像面积小、无形状及纹理特征、速度快等典型目标成像特性和技术难点，开展各种复杂环境下红外弱小目标检测新理论和新方法研究，以期提高空间光电探测系统目标探测、识别与跟踪性能。主要研究内容包括：（1）红外成像特性分析及目标正演模拟；（2）红外图像增强及高分辨重建技术；（3）红外弱小目标稀疏动态反演及检测技术；（4）数据测试与算法性能评价等。.课题组成员经4年多的技术攻关和探索研究，取得了一系列创新研究成果。（1）提出了基于“谱”反演、四方向全变分及高阶交叠组稀疏（OGS）的红外图像增强及超分辨率重建方法；（2）依据背景低秩、目标稀疏的红外图像模型和特点，提出了非负约束变分模态分解、稳定多子空间学习（SMSL）及稳定主成分追踪（SPCP）等3种红外弱小目标的稀疏动态反演检测算法，减低了虚警率；（3）为了减少目标检测的多解性、进一步提高检测性能，提出了加权截断核范数、非凸秩近似极小化联合l2,1范数及Lp-核范数联合正则化的红外弱小目标检测方法。（4）将经典矩阵低秩与稀疏分解模型推广到张量恢复模型，提出了基于张量鲁棒主成分分析的红外弱小目标检测方法，极大地提高了红外目标检测的运算效率。.通过探索和深入研究，形成了红外成像特性分析及建模、目标正演、稀疏动态反演及目标检测等完整的红外弱小目标检测方法和技术系列。取得的研究成果在对地观测中的虚警源及目标检测等光电探测领域进行了大量仿真和实测数据的测试评价，验证了本课题研究成果的有效性和可行性，达到了预期目标和技术指标要求，在军事、安防及遥感领域具有良好的应用前景。