Hyperspectral image denoising via spectral noise distribution bootstrap

Abstract

Hyperspectral image (HSI) denoising is an ill-posed problem, leading to integrating proper prior knowledge about hyperspectral noise is critical to developing an efficient denoising method. Most existing methods share a common assumption that all bands have equal noise intensity. However, such assumption runs counter to the practical HSIs, leading to unpleasant denoising results. To tackle this, we intend to investigate the intrinsic properties of real HSI noise in the spectral dimension and construct a novel denoising framework bootstrapping by spectral noise distribution N^, termed N^-Net. On the one hand, we develop dense and sparse recurrent calculations, exploiting intrinsic properties of HSI noise (i.e., diversity, dense dependency, and global sparsity) to estimate spectral noise distribution. On the other hand, having the estimated spectral noise distribution, we develop a bootstrap mechanism with a repetitive emphasis on its guidance for subsequent spatial noise separation and clean HSI recovery, ensuring a more delicate denoising effect. In particular, we verify that the proposed denoising framework can achieve promising denoising performances due to the merit of spectral noise distribution bootstrapping, which also promotes new insights for future related research. The code is avaliable at https://github.com/EtPan/N-Net.