SSIT: A spatial–spectral interactive transformer for hyperspectral image denoising

Convolutional neural networks have been successfully applied to hyperspectral image denoising, but they cannot effectively capture global information in the image. To address this issue, spatial–spectral transformers have recently been proposed to capture global spectral correlation and non-local spatial similarity. Nevertheless, these solutions cannot capture essential spatial–spectral characteristics in hyperspectral images well and fuse spatial and spectral features effectively. Based on this, we propose a spatial–spectral interactive transformer for hyperspectral image denoising. This method comprises a spatial–spectral attention module and a spatial–spectral interactive module. The former comprises a spatial permuted self-attention to model local and non-local spatial similarities and a spectral compression attention to extract spectral features with global spectral correlation and low-rank property. The latter further integrates the spatial–spectral features interactively. Additionally, we design a spectral-split feed-forward network to improve the model’s capability to model spatial–spectral characteristics. Extensive experiments on synthetic and real noisy hyperspectral images demonstrate the effectiveness of the proposed method. The reproducible code is available at https://github.com/jxnuczy/SSIT . • A spatial–spectral interactive transformer (SSIT) is proposed for HSI denoising. • Dynamic spatial–spectral interaction fuses spatial–spectral features better. • Local and non-local spatial similarities of HSI can be captured simultaneously.