高光谱图像有着比常规RGB图像更多的光谱带，包含了更加丰富的信息，在遥感监测、地形勘探等任务上有着十分重要的应用。高光谱图像的存储占用大，传输成本高，通常情况下会采用一种快照压缩成像系统将高光谱图像压缩成一张二维的快照估计图来进行存储和传输。当需要用到光谱图像来分析时，再从压缩的估计图进行重建。现阶段，先进的高光谱图像重建技术仅由少数欧美国家掌握，我国该领域的研究相对落后。因此，如何鲁棒、准确、高效地重建出原始的高光谱图像便成为了一个亟待解决的重要难题。 首先，本文分析了传统高光谱图像重建方法的泛化能力差的弊端，同时指出当前深度学习领域基于卷积神经网络的方法在捕获长程依赖关系的不足。针对这两个问题，本文设计了一种编码孔径掩膜引导的光谱维度Transformer算法。该算法在使用不到其他算法5%参数量和计算量的前提下，性能高出了2.55 dB。此外，该算法还在CVPR 2022上举办的NTIRE高光谱图像重建竞赛中斩获冠军。 其次，本文针对光谱特征在空间维度上分布稀疏而引发的计算效率低下的问题，提出一种多尺度哈希映射光谱表征聚合算法。该算法可将空间分布稀疏分散的光谱表征定位筛选出来，然后将计算量集中部署在光谱信息密集的区域，在使用更低计算代价的同时，性能上提升了0.68 dB。 接着，本文针对了以往深度展开式算法无法感知快照压缩成像系统的退化模式与病态度的问题，设计了一种退化模式可感知的深度展开式算法。该算法能从成像系统中估计缩放系数与噪声水平两个重要的信息参数用以指导后续的迭代学习过程，不仅具有可解释性与理论保障，同时在不增加额外计算负担的前提下，性能上也超过了当前最好的深度展开式算法4 dB。 最后，本文针对真实成像场景下噪声会干扰光谱图像重建效果的问题，提出了一种像素级噪声可感知的生成式对抗训练算法来模拟噪声的分布并去除噪声，使上述高光谱图像重建算法免于噪声的干扰，更稳定地在真实场景中复原光谱图像。
Compared to normal RGB images, hyperspectral images have more spectral bands to store richer information of the captured scenes. Thus, hyperspectral images have wide and important applications like remote sensing and terrain exploration. Storing and transmitting hyperspectral images cost a lot. Snapshot compressive imaging systems are usually employed to compress the 3D hyperspectral image data cube into a 2D measurement for low-cost storage and transmission. When using hyperspectral images, they are reconstructed from the compressed measurement. Currently, the advanced hyperspectral image reconstruction techniques are mastered by a few European and American countries. The research of this topic in China is relatively backward. How to robustly, accurately, and efficiently reconstruct hyperspectral images still remains under-explored.Firstly, this thesis analyzes that traditional model-based algorithms suffer from poor generalization ability and existing deep learning methods rely on convolutional neural networks, showing limitations in capturing long-range dependencies. To tackle these issues, this thesis proposes a mask-guided spectral-wise Transformer algorithm. This method outperforms previous methods by 2.55 dB while costing 5% Params and FLOPs. Besides, this algorithm won the first place in the CVPR 2022 NTIRE Spectral Recovery Challenge.Secondly, to address the low-efficiency problem caused by the spatial sparsity property of hyperspectral images, this thesis proposes a coarse-to-fine hash mapping spectrum aggregation algorithm. This algorithm can detect regions with dense spectral signals, screen them out, and focus the computation on these regions. It achieves 0.68 dB improvements while requiring less computational costs.Thirdly, to handle the issue that previous deep unfolding methods do not perceive the degradation patterns and ill-posedness degree of the snapshot compressive imaging system, this thesis proposes a degradation-aware deep unfolding mathematical algorithm framework. This framework can estimate two important informative parameters, scaling factor and noise level, to direct the subsequent iterative learning. This algorithm not only has interpretability and theoretically proven properties but also outperforms state-of-the-art deep unfolding methods by 4 dB, without additional computational burden.Finally, the measurements of real scenes captured by snapshot compressive imaging systems are noisy, which has negative effects on the reconstruction performance. To address this issue, this thesis designs a pixel-level noise-aware generative adversarial training algorithm to model the noise distribution and then remove the noise. With the proposed techniques, hyperspectral image reconstruction can be free from noise perturbation and applied in real imaging scenes more robustly.
