A Spatial-Spectral Transformer Network With Total Variation Loss for Hyperspectral Image Denoising

Hyperspectral image (HSI) denoising has an essential effect on HSI analysis and interpretation. In recent years, denoising methods based on convolutional neural networks (CNNs) have made great progress. However, the convolution kernel in the CNN model is content-independent, and the ability to capture long-distance correlation is weak, which leads to spectral distortion and edge blurring. To address this problem, we propose a spatial-spectral transformer network for HSI denoising, which introduces the shifted window-based transformer method to denoise HSIs by modeling image content correlation while preserving the local inductive bias. In detail, to jointly explore the spatial-spectral features, we first formulate the spatial-spectral cubes as network input, which are composed of the current band and its adjacent fixed K-bands. Second, these spatial-spectral cubes are forwarded to cascaded hyperspectral transformer blocks (HTBs) with skip connection for deep feature extraction. The HTB contains multiple transformer layers based on different window partitioning, which not only reduces memory cost, but also enhances the feature extraction capability. Finally, for the denoised B spatial-spectral cubes, we average the pixels of overlapping spectral bands to generate a complete HSI. Furthermore, we introduce the total variation (TV) to preserve the smoothness structures of HSIs. The experimental results on simulated and real data indicate that the proposed spatial-spectral transformer denoising (SSTD) is superior to other mainstream learning-based HSI denoising algorithms.