HYPERSPECTRAL IMAGE DENOISING USING WAVELET SPATIAL-SPECTRAL ATTENTION

Hyperspectral image (HSI) denoising can produce higher-quality data, providing a solid foundation for subsequent applications. However, existing methods often ignore the preservation of the high-frequency information during the denoising process, resulting in poor image details that significantly hinder the performance of subsequent tasks such as classification. To address these limitations, a Wavelet Spatial-Spectral Attention Block (WSSB) based method is proposed in this paper. Specifically, the designed WSSB first obtains the low-frequency and high-frequency components of the input HSI by tensor wavelet transform. Since noise is usually contained in the high frequency component along with structural information, WSSB can effectively remove noise in high frequency component and enhance features in low frequency components by applying spatial attention to the low frequency component and group spectral attention to the high-frequency component, respectively. Then, the high frequency features after noise removal are combined with the low frequency information to reconstruct the image. Finally, the high frequency features affected by denoising can be restored with the assistance of enhanced low frequency information via a local-global aware block consist of Swin Transformers and residual convolution. Extensive experiments on various datasets have demonstrated the effectiveness of the proposed method.