基于深度学习的高光谱图像异常目标检测算法研究

Deep learning methods can effectively extract high-level abstract features of hyperspectral images. Researches are carried on the deep learning theory to anomaly target detection for hyperspectral images. The deep learning model based on joint representation is first proposed. In the original dataset and its sparse dataset, the joint-representation DBN model is conducted on the error between the test point vector and its DBN reconstruction. The joint-representation tensor-SAE model is conducted on the differences between the test tensor and its projection onto the background dictionary tensor subspace. The joint-representation tensor-CNN model is conducted on the differences between the test tensor and its tensor-CNN convolution with the background tensors. The adaptive deep learning model is then proposed. The adaptive DBN model is conducted on the differences between the test point vector and the dictionary vectors and the differences between the local background vectors and the dictionary vectors. The adaptive tensor-SAE model is conducted on the differences between the test tensor and its projection onto the dictionary tensor subspace and the differences between the background tensor blocks and their projections onto the dictionary tensor subspace. The adaptive tensor-CNN model is conducted on the differences between the test tensor and its tensor-CNN convolution with the dictionary tensor blocks and the differences between the local background tensor blocks and their tensor-CNN convolutions with the dictionary tensor blocks. The success of this project will be of great significance as well as application value to the effective exploration and utilization of hyperspectral images.

深度学习方法能有效地提取高光谱图像的高层次抽象特征，课题围绕深度学习理论深入研究异常目标检测的热点技术。提出联合表示深度学习模型，分别在原始高光谱数据和其稀疏矩阵数据中，利用测试点向量和其DBN模型重建之误差建立联合表示DBN模型；利用对测试点张量和其在背景张量的映射之差建立联合表示SAE模型；利用测试点张量和其与背景张量的张量-CNN卷积之差，建立联合表示张量-CNN模型。提出自适应深度学习模型，利用测试点向量和局部背景向量分别与字典向量之差建立自适应DBN模型；利用测试张量块和邻域张量块分别在背景字典张量中的映射之差建立自适应张量-SAE模型；利用测试张量块和邻域张量块分别与背景字典张量块的张量-CNN卷积之差建立自适应张量-CNN模型。本课题的成功研究对高光谱图像信息的有效挖掘和利用有着重要的理论意义和应用前景。

通过高光谱遥感图像，人类能够认知原来不能探知的物质或不能识别的目标。异常目标检测无需先验信息，是高光谱遥感图像的研究热点之一。本项目针对高光谱图像异常目标检测精度的提升，利用深度学习理论，开展了基于联合表示的深度学习的高光谱图像异常目标检测研究，提出了联合稀疏表示和改进张量映射的异常检测算法，联合栈式自动编码和分数傅里叶变换的异常检测算法以及基于分数傅里叶变换和深度置信网络的异常目标检测算法，从多角度采用深度学习模型，提升检测精度；开展了基于自适应深度学习的高光谱图像异常检测研究，提出了基于张量的改进卷积神经网络异常检测算法，基于张量的分数傅里叶变换和改进的卷积神经网络模型以及基于分数傅里叶变换的联合自适应子空间检测模型，避免了异常目标检测中未知地物光谱的限制；此外，结合深度学习、分数傅里叶变换以及计算机视觉，提出了基于分数傅里叶变换的张量RX的异常检测算法和基于改进的中心注意力网络的张量RX的异常检测算法，有效地提升了RX检测算法的精度。本项目取得的研究进展和成果已经发表在国内外学术刊物，包括8篇国际期刊和1篇国内期刊论文，已申请专利1项，出版专著1部。