多源遥感图像识别关键技术研究

The research in remote sensing image object recognition is a hotspot in multimedia analysis and understanding and pattern recognition, and it is of significant value in both theories and applications. The research is also of great value for solving the problem of high resolution remote sensing image understanding and processing. The project is proposed to focus on the key scientific problems and application requirements of recognition for multi-source remote sensing images using bridges, airports, roads and other typical objects. A registration algorithm between SAR and panchromatic images combined with Ratio gradient and cumulative residual entropy is presented to improve the accuracy and speed of registration. The project studies the visual recognition mechanism for multi-source remote sensing images, applies the principle of visual global precedence in cognitive science to object recognition, and establishes a target identification framework combined multi-source remote sensing image complementarity and multi-scale feature coupling. The project analyzes the target feature complementarity for multi-spectral, panchromatic and SAR images, and the relationship between the multi-scale features, studies the feature extraction technology and coupling mechanism for multi-source remote sensing images. A strong adaptability and high accuracy target recognition reasoning model is studied using image characteristics and experience knowledge. The research achievements will help to break through the current target recognition technology bottlenecks and effectively improve the recognition rate of typical targets, and can provide reliable theoretical basis and effective solution for related applications.

遥感图像目标识别技术是多媒体分析与理解、模式识别等领域的研究热点，具有重要的理论和应用价值，对解决高分辨率遥感图像的理解和处理问题具有重要意义。本项目针对多源遥感图像目标识别中的关键科学问题和应用需求，以桥梁、机场、道路等典型目标为研究对象，研究Ratio梯度与交叉累积剩余熵相结合的SAR与全色图像自动配准算法，提高配准精度和速度；研究针对多源遥感图像的视觉识别机理，将认知科学中的视觉整体性原理运用到目标识别中，建立一种融多源遥感图像互补、多尺度特征耦合的目标识别框架；分析目标特征在多光谱、全色、SAR图像中的互补作用，以及多尺度间的关系，研究多源遥感图像的特征提取技术与耦合机制；利用遥感图像特征和经验知识，研究适应性强、准确性高的目标识别推理模型。本项目的研究成果将有助于突破当前目标识别的技术瓶颈，有效提高典型目标的识别率，可以为相关应用提供可靠的理论基础和有效的解决途径。

遥感图像目标识别技术是多媒体分析与理解、模式识别、遥感等学科领域的研究热点，具有重要的理论和应用价值。本项目提出了一种二维希尔伯特变换和BEMD结合的图像去噪方法。由于光学图像与SAR图像成像机理不同，本项目提出了一种基于Ratio梯度和CCRE的SAR与全色图像快速配准算法，提高了配准速度和精度，提出了多光谱与全色图像的高精度配准算法。目前的识别方法仅侧重于研究目标的局部特征，忽略了目标的整体特征，本项目提出了一种基于模糊阈值的自适应图像分割方法，建立了基于视觉整体性的目标识别框架，研究设计了多源图像特征提取方法、面向目标的多源遥感图像特征耦合机制，研究实现了图像特征与知识结合的推理模型。提出了一种多源遥感图像互补特征和知识相结合的桥梁识别方法、基于视觉整体性的建筑物识别方法、遥感图像多尺度特征与知识结合的机场识别方法、多源遥感图像特征与知识结合的码头识别方法，以及基于改进BP神经网络的道路识别方法，有效提高了典型目标的正确识别率。本项目已经授权发明专利8项，受理发明专利4项，获软件著作权13项，发表SCI检索期刊论文4篇，EI检索期刊论文7篇，EI检索会议论文6篇，核心期刊论文11篇，普通期刊论文8篇，录用核心期刊论文4篇。项目负责人指导了硕士研究生16名，其中2名获校级优秀硕士学位论文。