时敏目标天/空在线检测与识别

Time-sensitive target is the important target that must be detected and recognized successfully within the limited time windows, and it is of great strategic significance for the military and civil applications to detect and recognize time-sensitive target online. However, aerospace online detection and recognition of time-sensitive target is very challenging due to the time-variation of time-sensitive target, the complexity of the volumetric multi-sensor data and the peculiarity of online application. To find the solutions to the bottleneck issues that largely hinder the practical applications, this project aims to explore new techniques such as sparse representation and dictionary learning to develop new theories and new methodologies for online detection and recognition of time-sensitive target. Specifically, the new detection-recognition framework for time-sensitive target will be proposed based on the cognition principles, and the key techniques will be studied in depth, such as the cognition-invariant feature representation of time-sensitive target on the multi-source images, the data-driven target-of-interest specific feature enhancement and suppression, the utilization of objectness-based feature representation and spatial-temporal salience or spatial-temporal change salience for detecting and recognizing time-sensitive target. By taking advantages of the proposed framework and techniques, new application fashion will be explored based on multisource multi-temporal signals or images and multi-scale/multi-level strategy. To demonstrate the effectiveness of the project, the prototype system will be developed based on the parallel implementation of software and DSPs. In summary, the project will provide the sound theoretical and technical foundation for the practical applications of aerospace online detection and recognition of time-sensitive target, and help accelerate the development of intelligent automatic target recognition in China.

时敏目标是指高价值、稍纵即逝的重要目标，时敏目标的在线检测和识别在军事、民用方面有着极其重要的应用价值。然而，由于时敏目标的时变性、海量多源数据的复杂性和在线应用的特殊性，时敏目标的天/空在线检测和识别极具挑战性。本项目旨在利用稀疏表示、字典学习等新技术研究时敏目标在线检测、识别的关键理论和关键技术。具体的，本项目将在认知机理的基础上研究时敏目标在线检测识别新框架；研究时敏目标在多源图像上的感知不变特征表示、由数据驱动的基于感兴趣目标类型的特征增强与抑制、基于“目标性”特征表示和时空显著性、时空变化显著性的时敏目标检测与识别；在此基础上，研究基于多源、多时相信号/图像、多尺度、多级别策略的时敏目标天/空在线检测与识别新应用模式，研制基于软件和多DSP并行的原型系统。本项目的成功实施，可为时敏目标的天空在/线检测和识别应用提供坚实的理论与技术支撑，推动我国智能目标识别技术的跨越式发展。

时敏目标是指高价值、稍纵即逝的重要目标，时敏目标的在线检测和识别在军事、民用方面有着极其重要的应用价值。然而，由于时敏目标的时变性、海量多源数据的复杂性和在线应用的特殊性，时敏目标的天/空在线检测和识别极具挑战性。本项目从认知机理、算法、应用模式及示范系统四个方面对上述难点进行了全面深入研究。首先，根据成像机理和认知机理的最新进展，提出了关系学习新框架，该框架不仅可以描述、反映时敏目标在线检测和识别的关键难点，还将目标识别、目标时空变化分析、目标上下文识别等若干步骤和技术纳入统一的框架。其次，基于关系学习理论和深度学习，提出了一系列新的有效算法，包括基于kNN哈希的目标识别、基于任务自适应性的目标在线检测、基于时相变化的变化目标检测与识别、基于深度关系学习的目标识别、目标跟踪及目标时敏性分析、目标变化区域与变化类型识别、深度特征学习与目标变化分析、基于自适应图割和多特征的道路提取、基于信息熵的卷积神经网络的地物分割(精识别)算法、基于判别分析和鲁棒回归的半监督高光谱图像分类等。然后，研究了空间信息网络下的时敏目标的天/空在线检测和识别新应用模式，通过哈希学习及哈希保序约束实现在线时敏目标快速检测和识别，通过基于目标重要性的自适应快速采样实现在线特征快速压缩，从而保证了在计算资源、通信资源、存储资源受限情况下的目标快速识别和传输。最后，开发了时敏目标在线检测和识别示范系统，可以在单时相图像上识别时敏目标，也可以在多时相图像上分析时敏目标的时空变化，部分算法用DSP和FPGA实现。