基于深度学习的高光谱图像分辨率增强与目标检测联合优化方法

Target detection from hyperspectral imagery is a hot topic in the field of remote sensing information processing. However, due to the low spatial resolution, mixed pixels, blurring contour of the target, and separation mechanism between super-resolution task and target detection task in hyperspectral imagery, it is difficult to ensure the accuracy of target detection. In order to further improve the detection accuracy, this project mainly focuses on deep learning based jointly optimization method for super-resolution and detection of hyperspectral imagery. The main contributions are as follows. Firstly, the high-level features of the spatial-spectral information in hyperspectral imagery are extracted by deep neural networks, and the super-resolution task and target detection task are carried using the extracted high-level features. Moreover, by incorporating “spectral fidelity” and “spectral identifiability” into optimizing objective function, a deep learning based multi-task driven jointly optimization model is proposed. Finally, in order to efficiently and effectively solve the optimizing model, we will design a paralleled algorithm. The project not only provides powerful technical support for hyperspectral imagery super-resolution task and target detection task, but also provides a new method and pattern for wide application of hyperspectral remote sensing in earth observation, deep space exploration and other related fields.

高光谱图像目标检测是遥感信息处理领域的研究热点。然而，高光谱图像空间分辨率低，存在混合像元和目标轮廓信息模糊等缺点，以及空间分辨率增强任务与目标检测任务目前所采取的分离机制，这些都会影响目标检测的精度。为了提高目标检测的精度，本项目主要研究基于深度学习的高光谱图像分辨率增强与目标检测联合优化模型与算法。主要内容包括：高光谱图像空间结构和光谱信息的深度特征学习方法，以及利用提取的空谱特征进行空间分辨率增强和目标检测；以分辨率增强的“光谱保真性”和目标检测的“光谱可鉴别性”为优化目标，提出多任务驱动的基于深度学习的联合优化模型，并设计出高效的并行求解算法。本项目不仅为高光谱遥感分辨率增强和目标检测提供强有力的技术支撑，同时为高光谱遥感在对地观测和深空观测等相关领域的实际应用提供新的方法和手段。

项目以高光谱图像分辨率增强和检测为研究对象，利用深度学习理论探索高光谱图像空谱特征提取、分辨率增强和检测等问题的应用效果。首先，针对高光谱图像重构问题，利用深度学习理论提取高光谱图像的空谱特征，提取了一种基于空谱联合的密集残差网络的高光谱图像重构方法。其次，为了增强遥感图像的分辨率，提出了一种基于多方向子带深度神经网络的遥感图像融合方法。该方法深入挖掘图像几何结构的方向性信息和光谱波段之间的相关性信息，将其融入到深度神经网络融合模型中，使融合后的图像具有锐利的边缘结构和丰富的纹理信息。此外，利用密连网络共享特征、任意层间互联的特性，通过改进密集连接模块和引入瓶颈层，提出了一种基于新型密连卷积神经网络架构的遥感图像融合方法。最后，针对高光谱图像分类检测问题，利用局部二值模式来提取超像素内空间信息，结合高光谱图像空谱先验信息，提出了一种基于局部二值模式和空谱多核高光谱图像分类方法，该方法能有效地防止超像素内边缘特征的遗失。通过本项目的研究，进一步加强了深度学习理论深度和应用广度，促进了遥感图像在智能识别和精确解译方面的发展，为遥感图像的分类和检测奠定了坚实基础。