[论文解读] Degradation-Noise-Aware Deep Unfolding Transformer for Hyperspectral Image Denoising

Haijin Zeng, Jiezhang Cao+5 more|arXiv (Cornell University)|May 6, 2023

Image and Signal Denoising Methods被引用 5

一句话总结

本文提出DNA-Net，一种新型的降质-噪声感知深度展开Transformer网络，用于高光谱图像（HSI）去噪，能够联合建模稀疏噪声与高斯噪声，并在迭代过程中估计降质参数。通过集成U-结构局部-非局部-光谱注意力（U-LNSA）模块，DNA-Net能够捕捉长程依赖关系与光谱相关性，在计算成本更低的同时，显著提升了在强噪声条件下的鲁棒性与性能表现。

ABSTRACT

Hyperspectral imaging (HI) has emerged as a powerful tool in diverse fields such as medical diagnosis, industrial inspection, and agriculture, owing to its ability to detect subtle differences in physical properties through high spectral resolution. However, hyperspectral images (HSIs) are often quite noisy because of narrow band spectral filtering. To reduce the noise in HSI data cubes, both model-driven and learning-based denoising algorithms have been proposed. However, model-based approaches rely on hand-crafted priors and hyperparameters, while learning-based methods are incapable of estimating the inherent degradation patterns and noise distributions in the imaging procedure, which could inform supervised learning. Secondly, learning-based algorithms predominantly rely on CNN and fail to capture long-range dependencies, resulting in limited interpretability. This paper proposes a Degradation-Noise-Aware Unfolding Network (DNA-Net) that addresses these issues. Firstly, DNA-Net models sparse noise, Gaussian noise, and explicitly represent image prior using transformer. Then the model is unfolded into an end-to-end network, the hyperparameters within the model are estimated from the noisy HSI and degradation model and utilizes them to control each iteration. Additionally, we introduce a novel U-Shaped Local-Non-local-Spectral Transformer (U-LNSA) that captures spectral correlation, local contents, and non-local dependencies simultaneously. By integrating U-LNSA into DNA-Net, we present the first Transformer-based deep unfolding HSI denoising method. Experimental results show that DNA-Net outperforms state-of-the-art methods, and the modeling of noise distributions helps in cases with heavy noise.

研究动机与目标

• 解决现有基于学习的HSI去噪方法中缺乏噪声分布建模的问题。
• 克服基于CNN的模型在捕捉长程空间与光谱依赖关系方面的局限性。
• 通过将降质感知参数估计集成到深度展开框架中，提升HSI去噪的可解释性与性能表现。
• 构建一种混合模型，融合基于模型的方法与深度学习的优势，实现鲁棒的HSI复原。

提出的方法

• 提出一种基于最大后验概率（MAP）的深度展开框架（DNA-Net），在优化过程中显式建模稀疏噪声与高斯噪声。
• 直接从含噪HSI与降质模型中估计超参数与噪声特性，以指导每一级展开迭代。
• 引入一种U-结构局部-非局部-光谱注意力（U-LNSA）模块，以联合捕捉局部特征、光谱相关性与非局部依赖关系。
• 将优化过程展开为端到端可训练的神经网络，每一阶段均结合保真度项与先验学习，并通过注意力机制进行特征精炼。
• 采用基于Transformer的架构替代传统CNN，实现全局上下文建模与更优的长程特征学习能力。
• 在每一迭代中集成噪声感知损失与参数估计，以增强在不同噪声条件下的鲁棒性。

实验结果

研究问题

• RQ1 深度展开框架是否能有效建模HSI去噪中稀疏与高斯噪声分布？
• RQ2 引入降质感知参数估计在多大程度上提升了去噪性能与泛化能力？
• RQ3 自注意力机制如U-LNSA在捕捉HSI数据中的长程空间与光谱依赖关系方面，相较于CNN的性能优势有多大？
• RQ4 将基于Transformer的架构整合到深度展开框架中，是否能相比现有SOTA方法实现更优性能与更低计算成本？
• RQ5 在所提出的DNA-Net架构中，迭代次数与性能之间的权衡关系如何？

主要发现

• 在S1数据集的Case -3噪声条件下，DNA-Net的PSNR达到44.168 dB，较先前SOTA方法SERT提升1.78 dB。
• 在S9数据集的Case -4噪声条件下，DNA-Net-B的PSNR达到43.780 dB，较SERT高出1.79 dB，展现出对强噪声更强的鲁棒性。
• 所提出的U-LNSA模块显著提升性能：引入光谱与非局部注意力后，PSNR相比仅使用局部注意力的模型提升超过1.5 dB。
• DNA-Net-B在5次迭代下获得的PSNR与SSIM均高于DNA-Net-L在2次迭代下的表现，尽管推理时间有所增加。
• DNA-Net在实现SOTA性能的同时，显著降低了计算成本与显存占用，优于以往方法。
• 消融实验表明，噪声感知参数估计与多头注意力机制对性能提升至关重要，尤其在高噪声环境下表现更显著。