Abstract: To address the insufficient exploitation of spectral priors, inadequate modeling of spectral variations at different scales, and limited spatial detail recovery in existing hyperspectral image denoising methods, a hyperspectral image denoising method integrating spectral prior and multiscale feature enhancement was proposed. A hierarchical encoder-decoder architecture was adopted to extract multiscale spatial-spectral features. A distance prior among spectral-related latent feature dimensions was introduced into the attention computation to adaptively calibrate spectral feature interactions, thereby enhancing the modeling of spectral continuity structure and long-range correlations. Meanwhile, multiscale spectral convolution, adaptive spatial gating, and high-frequency detail enhancement were combined to recover local spectral variations, edge textures, and spatial structures. Furthermore, a joint loss function consisting of Charbonnier reconstruction loss, spectral consistency loss, and spatial gradient loss was constructed to constrain pixel-level reconstruction accuracy, spectral directional consistency, and spatial structure preservation capability. On the Harvard dataset, the proposed method achieved PSNR values of 40.62, 38.41, and 36.90 dB under Gaussian noise with standard deviations of 30, 50, and 70, respectively, outperforming the best corresponding comparison model Restormer by 1.27, 1.22, and 1.27 dB. Under Poisson-Gaussian mixed noise, the PSNR reached 42.66 dB, compared with SSIT, an improvement of 0.88 dB. On the Realistic Data, the PSNR, SSIM, SAM, and SPR reached 30.74 dB, 0.9227, 2.49°, and 0.9982, respectively. In the ablation experiment, the complete model improved the PSNR by 0.61 dB and reduced the SAM by 1.47° compared with the baseline model. The proposed model reduces pixel reconstruction errors and spectral deviations under different noise conditions while improving the preservation of edge textures and spatial structures.

Key words: hyperspectral image denoising, transformer, attention mechanism, multiscale feature fusion, spatial-spectral modeling

摘要： 针对现有高光谱图像去噪方法对光谱连续先验利用不足、不同尺度谱域变化建模不充分以及空间细节恢复能力有限的问题，提出一种融合光谱先验与多尺度特征增强的高光谱图像去噪网络。该网络采用层次化编码器-解码器结构提取多尺度空间-光谱特征，在注意力计算中引入光谱相关潜在特征维度的距离先验，对谱域特征交互进行自适应校准，以增强光谱连续结构和长距离相关性建模能力；通过多尺度谱域卷积、自适应空间门控和高频细节增强机制，联合恢复局部谱域变化、边缘纹理和空间结构。进一步构建由Charbonnier重建损失、光谱一致性损失和空间梯度损失组成的联合损失函数，约束像素重建精度、光谱方向一致性和空间结构保持。在Harvard数据集上，高斯噪声标准差为30、50和70时，峰值信噪比分别达到40.62、38.41和36.90 dB，较相应最优对比模型Restormer分别提高1.27、1.22和1.27 dB；在泊松-高斯混合噪声条件下，峰值信噪比达到42.66 dB，较SSIT提高了0.88 dB。在Realistic Data数据集上，峰值信噪比、结构相似性、光谱角和光谱保持率分别达到30.74 dB、0.9227、2.49°和0.9982。消融实验中，完整模型较基础模型的峰值信噪比提高0.61 dB，光谱角降低1.47°。结果表明，该模型能够降低不同噪声条件下的像素重建误差和光谱偏移，并改善边缘纹理和空间结构保持能力。

关键词: 高光谱图像去噪, Transformer, 注意力机制, 多尺度特征融合, 空间-光谱建模