Noise Intensity Guided Wavelet Domain Denoising of Hyperspectral Images

如何在不破坏边缘细节结构的情况下去除各种噪声，是高光谱图像(HSI)要解决的问题。文章提出了一种噪声强度指导小波域网络，即NGWDNet，它应用学习得到的噪声特征图来指导小波变换后的频域特征，从而完全去除高光谱图像的噪声。首先，通过设计的噪声估计模块(NEB)精准计算出各波段的噪声特征图。随后利用离散小波变换(DWT)将原始HSI分离为两个高频分支与一个低频分支。针对不同频率分支，NGWDNet采用差异化处理策略。对于富含纹理和细节特征的高频分支，将噪声图与之融合，并设计基于移位/窗口的多头自注意力融合模块(W/SW-MSAF Block)，强化对细节信息的捕捉与处理。对于具有较多平滑结构特征的低频分支，则直接应用3D和2D Unit，高效保留结构信息。同时，充分考量不同分支间的特征交互，借助高频特征引导低频分支精细结构的重建，进一步挖掘纹理细节特征。最后，部署了空间–光谱残差块(SSRB)，在融合双域的同时进一步探索空间–光谱相关性，以获得更好的去噪输出。在公开数据集上进行了定性和定量实验，结果表明，NGWDNet在可视化和客观评价指标方面优于最先进的方法。How to remove all kinds of noise without destroying the edge detail structure is a problem to be solved in hyperspectral images (HSI). This paper proposes a noise intensity guided wavelet domain network, namely NGWDNet, which uses the learned noise feature map to guide the frequency domain features after wavelet transform so as to completely remove the noise of the hyperspectral image. Firstly, the noise characteristic map of each band is accurately calculated by the designed noise estimation module (NEB). Then, the original HSI is separated into two high-frequency branches and one low-frequency branch by discrete wavelet transform (DWT). For different frequency branches, NGWDNet adopts a differentiated processing strategy. For high-frequency branches rich in texture and detail features, the noise map is fused with it, and a multi-head self-attention fusion module based on shift/window (W/SW-MSAF Block) is designed to enhance the capture and processing of detail information. For low-frequency branches with smoother structural features, 3D and 2D Units are directly applied to efficiently retain structural information. At the same time, the feature interaction between different branches is fully considered, and the high-frequency features are used to guide the reconstruction of the fine structure of the low-frequency branches to further explore the texture detail features. Finally, we deployed the spatial-spectral residual block (SSRB) to further explore the spatial-spectral correlation while fusing the dual-domain to obtain better denoising output. We conducted qualitative and quantitative experiments on public datasets. The results show that NGWDNet is superior to the most advanced methods in terms of visualization and objective evaluation indicators.