融合频域调制的多尺度高光谱图像去噪网络

Multi-Scale Hyperspectral Image Denoising Network with Frequency-Prior Guided Modulation

摘要: 高光谱图像(Hyperspectral Image, HSI)在遥感解译、目标识别和精细分类等任务中具有重要应用价值。然而，受传感器噪声和复杂成像环境影响，高光谱图像易受到高斯噪声干扰导致结构退化和光谱失真。针对现有方法在细节恢复与光谱一致性保持方面的不足，本文提出一种融合频域先验引导与特征调制机制的高光谱图像去噪网络。该方法构建多尺度特征提取网络，将频域统计信息嵌入特征调制过程，并设计三参数调制模块，实现对不同尺度特征响应的自适应调控。同时，引入轻量化通道增强模块和弱高频残差增强模块，以提升边缘纹理恢复能力；并在损失函数中加入光谱角约束，以增强光谱一致性。ICVL数据集上的实验结果表明，本文方法在PSNR、SSIM和SAM等指标上均取得较优性能，可视化结果进一步验证了其在边缘结构恢复与细节重建方面的有效性。

Abstract: Hyperspectral images (HSIs) play an important role in remote sensing interpretation, target recognition, and fine-grained classification. However, due to sensor noise and complex imaging conditions, HSIs are often contaminated by Gaussian noise, stripe noise, and mixed noise, resulting in spatial structure degradation and spectral distortion. To address the limitations of existing methods in detail restoration and spectral consistency preservation, this paper proposes a hyperspectral image denoising method that integrates frequency-prior guidance and feature modulation. The proposed method constructs a multi-scale feature extraction network, embeds frequency-domain statistical information into the feature modulation process, and designs a triple-parameter modulation module to adaptively regulate feature responses at different scales. Meanwhile, a lightweight channel enhancement module and a weak high-frequency residual enhancement module are introduced to improve edge and texture restoration. In addition, a spectral-angle constraint is incorporated into the loss function to enhance spectral consistency. Experimental results on the ICVL dataset show that the proposed method achieves superior performance in terms of PSNR, SSIM, and SAM. Visual results further verify its effectiveness in edge structure recovery and detail reconstruction.