Hyperspectral Denoising Based on Latent Noise Information Extraction

In the field of denoising hyperspectral images, the existing studies generally focus on utilizing highly similar and relevant spectral information. However, they overlook the significance of modelling noise, making it difficult to effectively extract noise information and achieve optimal denoising performance. In this paper, we introduce an innovative network, named Latent Noise Modelling Network (LNE-Net), designed to effectively denoise hyperspectral images (HSI). This method consists of three parts: latent noise extraction (LNE), gate-based noise removal (GBNR), and spatial-channel pixel restoration (SCPR). Specifically, we first utilize Transformer to extract latent noise information. Subsequently, we use this latent noise information to generate a noise gating map, which is then employed to remove pixel information containing noise. Finally, we further adopt the multi-head self-attention mechanism to explore single-channel long-range spatial information and multi-channel global contextual information, and restore pixel information lost during the denoising process. Additionally, by introducing the top-k selection strategy, we effectively prevent additional noise interference during the restoration process, and further enhancing the restoration effect. The experimental results on Gaussian noise and complex noise in the ICVL dataset show that our LNE-Net performs well in denoising hyperspectral images and surpasses the existing state-of-the-art methods.