基于空间-谱间自适应稀疏表示的高光谱压缩成像方法研究

Hyperspectral imaging(HSI) has been widely applied in military and enviromental resource monitoring. The compressed sensing(CS) theory provides new opportunities to perform undersampling in data acquistion thus reduce the amount of the acquired data. Signals will be recovered by solving nonlinear optimization problems at the decoder side. The basic assumption of CS is that signals are sparse and sparser.representations usually lead to lower reconstruction error. In this project,we will explore the spatial-spectral sparsity and low rankness of HSI data. With the non-convex optimization methods in sparse and low rank signal reconstruction, we will propose the jointly and adaptively spatial-spectral sparse reconstruction models and improve the image resolution from highly undersampled data. Research work in the project will be conducted as follows: 1) Design low-complexity adaptively parametric sparsifying transforms to improve the sparsity of hyperspectral imaging data; 2) Propose fast dictionary learning methods to represent the prior structured features of HSI data; 3) Set up new low rank models of HSI data and propose new models combing low rank and sparsity; 4) Derive new non-convex and fast optimization reconstruction algorithms to further improve the quality of reconstructed signals. This project is expected to develop new fast HSI theories and sparsity-based image reconstruction methods, resulting in greatly improved HSI for many applications.

高光谱成像在军事和环境资源监测等领域有着重要应用。压缩感知可减少采样数据量重建光谱数据，并在地面解码端进行非线性解码，这种编码和解码的非对称性非常适用于高光谱成像。压缩感知理论要求信号是稀疏的，信号越稀疏信号重建误差越小。本项目拟从高光谱数据的空间-谱间的联合稀疏表示和低秩出发，利用压缩感知非凸重建算法，建立自适应联合稀疏重建模型，尽可能缩短成像时间的前提下提高分辨率。研究内容包括：1) 设计低复杂度的自适应参数化稀疏变换，提高图像的稀疏表示能力；2)建立谱间结构先验的快速字典训练方法，提高字典重建的速度；3) 提出高光谱图像新的低秩模型，并结合稀疏性进行光谱重建；4）推导非凸的稀疏和低秩重建算法，进一步提高图像的重建质量。本项目的实施可以促进高光谱成像理论的发展，增强高光谱成像在科学研究中的应用。

高光谱成像在军事和环境资源监测等领域有着重要应用。本项目从高光谱数据的空间-谱间的联合稀疏表示和低秩出发，利用压缩感知非凸重建算法，建立自适应联合稀疏重建模型，尽可能缩短成像时间的前提下提高分辨率。研究内容包括1、研究了压缩感知和深度学习理论下高光谱图像的分类算法，通过研究区域划分的联合稀疏表示分类方法和基于多尺度ResNet的空谱信息高光谱图像分类算法进一步提高高光谱图像的分类精度。2、项目组利用低秩Hankel矩阵进行波谱重建，得到了快速且高质量的重建算法；3、项目组提出一种改进SURF的遥感图像目标识别算法和基于卷积神经网络的遥感图像道路检测算法，提高图像目标识别精度。本项目的实施可以促进高光谱成像理论的发展，增强高光谱成像在科学研究中的应用。