基于稀疏观测系统的光场重构与目标识别方法研究

Object recognition has been widely used in many fields, and the accurate target recognition based on light field feature library is a recognized problem. During light field feature database refactoring and matching process, there are uncertainty interference factors, which there are the complicated relationship among. In addition, the target image information is incomplete, inconsistency, and there is the nonlinear dynamic relation between the known image and the unknown image. These factors cause the relation dynamic and uncertainty between multi-angle field feature database and the target image matching of, making the target recognition accuracy low. Therefore, combining with the object of a class of ground moving targets, this project, with the target recognition based on light field feature library as the background, shall conduct the following research: multi-angle light field characteristic library distributed space-time cooperative reconstruction method, the collaborative real-time data acquisition method under multi-angle light field state, and light field characteristic library perspective adaptive coordinated matching method under multiple constraints. This project will focus on the solution to the key scientific problems in the process of establishing the above methods: nonlinear time-varying relationship between various influencing factors analysis and modeling, and the construction of the single light field matching law linkage model. The research achievements of this project could provide theoretical and technical base for the application of the target recognition system based on light field feature library, thus, it is of great significance to upgrade the light field reconstruction and target recognition technology.

目标识别已在很多领域得到广泛应用，而基于光场特征库的目标精确识别却是公认难题。由于光场特征库重构与匹配过程中受到各种不确定性因素干扰，且干扰因素间关系错综复杂，加之目标图像信息的不完备性、不一致性，已知图像与未获取图像之间的非线性动态关系，使得多视角光场特征库与目标图像的匹配呈现动态性、不确定性，导致目标识别精度不高。为此，本项目拟以基于多视角光场特征库的目标识别为场景，以一类地面移动目标为研究对象，开展以下研究内容：多视角光场特征库分布式时空协同重构方法、多视角光场状态下协同实时数据采集方法和多约束条件下光场特征库多视角自适应协同匹配方法。重点解决上述研究内容中存在的一些关键科学问题：各影响因素间非线性时变关系分析及建模、各视角光场匹配律联动模型的构建。本项目的研究成果可为基于多视角光场特征库的目标识别系统的实用化提供理论方法和技术支持，对提高光场重构与目标识别技术具有重要意义。

目标识别已在很多领域得到广泛应用，而基于光场特征库的目标精确识别却是公认难题。针对光场特征库重构与匹配过程中存在的难题，项目负责人和项目组成员对国内外现有方法进行充分论证的基础上，开展了一系列光场重构与目标识别方面的研究，具体内容如下：1）针对光场的稀疏机制进行研究。将光场表示为不同视角下的子光场（即多视角光场），并进行稀疏表示与重构；从多智能体协同与协作角度出发，结合图理论和分布式协同理论，建立了区域全覆盖约束下多视角光场间协同与协作机制。2）针对稀疏光场重构方法进行研究：提出了基于多视角的光场重构方法，建立了区域全覆盖约束下多视角光场之间的协作机制；提出了基于GAN的多视角光场重构方法，建立了基于GAN的多生成器与多判别器模型来完成对光场的重构。3）针对遮挡情况下的光场重构与目标识别方法开展研究：提出了基于迁移强化学习的多视角光场重构方法，建立了相似度量模型，根据相似度阈值选择强化学习或特征迁移学习模型；建立了特征迁移学习模型，使用PCA得到源域与目标域最大特征嵌入空间并用于标签数据迁。相关研究成果已发表学术论文8篇，其中7篇被SCI检索；申请发明专利5件，授权2件。该项目研究内容及成果对复杂场景下的目标精确识别具有重要的理论指导意义和实际应用价值，并且该部分相关成已经进行实际推广应用。