面向高光谱图像的深度语义分类网络研究

Hyperspectral image has been successfully applied to different fields which include food science, earth science, and agricultural science due to its wide spectral range, high spectral density, and non-destructive scanning. As the core problem of hyperspectral image analysis, semantic classification for hyperspectral image has become hot topic at home and abroad. However, the currently existing semantic classification models for hyperspectral image perform not well enough in real-world applications due to insufficient labelled data, the spectral-spatial relationship, and the multiscale objects. This project based on the state-of-art deep learning frameworks, focuses on the research on deep semantic networks for hyperspectral image. Firstly, an unsupervised data argumentation strategy is proposed to avoid the over-fitting problem in deep networks. Then, a multi-layer fusion strategy is designed to avoid the over-segmentation problem or under-segmentation problem, and then a multi-dimensional and multi-layer convolution filter bank is used to represent spectral-spatial relationship. Finally, the output resolution is optimized to be as same as the input resolution. The proposed deep semantic network of this project can be applied in different scenarios, and thereby it will be a general framework for hyperspectral semantic classification. Besides, this project can lay a theoretical foundation of interdisciplinary research on hyperspectral image and deep learning, advancing the development of semantic classification for hyperspectral image in various research fields.

高光谱图像以其光谱范围广，光谱密度高，扫描无损性等优点，已经成功应用于食品科学、地球科学、农业科学等领域。作为高光谱图像分析的核心问题，高光谱图像语义分类已经成为国内外的研究热点，然而，由于受到标签数据不足，光谱空间关系和对象多尺度等限制，现有的面向高光谱图像的语义分类模型在实际应用中效果不佳。因此，本课题以现有的深度学习框架为基础，开展了面向高光谱图像的深度语义分类网络研究：首先提出无监督的数据扩充策略，避免了过拟合问题；接着设计多层融合策略，避免了过分割或欠分割现象；然后设计多维分层的卷积组，表示了高光谱图像的光谱空间关系；最后，对网络的输出精度进行提升，保证了输出输入分辨率相同。本课题研究的深度语义分类网络具有较强的通用性，可以适用于不同的高光谱图像应用场景。该课题的实施可为高光谱图像与深度学习的交叉发展提供理论依据，推动了高光谱图像语义分类在各个科学研究领域的应用发展。

高光谱图像以其光谱范围广，光谱密度高，扫描无损性等优点，已经成功应用于食品科学、地球科学、农业科学等领域。作为高光谱图像分析的核心问题，高光谱图像语义分类已经成为国内外的研究热点，然而，由于受到标签数据不足，光谱空间关系和对象多尺度等限制，现有的面向高光谱图像的语义分类模型在实际应用中效果不佳。因此，本课题以现有的深度学习框架为基础，研究了基于光谱注意力的编码解码高光谱语义分类网络：首先提出无监督和有监督的的数据扩充策略，避免了过拟合问题；接着设计面向光谱维度的光谱注意力模型表示光谱空间关系。然后设计基于transformer的编码结构，对高光谱图像的图像特征进行表达；最后，对网络的输出精度进行提升，保证了输出输入分辨率相同。本课题研究的深度语义分类网络具有较强的通用性，在不同的数据集在不同的评价指标上，都优于其他比对算法。该网络可以适用于不同的高光谱图像应用场景。该课题的实施可为高光谱图像与深度学习的交叉发展提供理论依据，推动了高光谱图像语义分类在各个科学研究领域的应用发展。