Hyperspectral Image Denoising Based on Multi-Resolution Gated Network with Wavelet Transform

Abstract

Hyperspectral image denoising is an essential pre-processing task. In this paper, a multi-resolution gated network based on wavelet transform (WMRGNet) is proposed for removing mixed noise of hyperspectral images. Firstly, based on the fact that hyperspectral images have strong spectral correlation, a spatial-spectral information extraction module is designed to use the current noisy band and its adjacent bands as the input of WMRGNet. Secondly, aim to fully consider the spatial local and global information of hyperspectral images, a multi-resolution feature extraction module is proposed, applying the discrete wavelet transform to divide the resolution into four scales, and the residual blocks to extract information of different resolutions. In addition, a gated layer is introduced for cross-resolution information interaction to enhance the feature fusion. Finally, a high-resolution image reconstruction module with multiple residual blocks is employed to extract high-resolution features. In the simulated data set experiments, WMRGNet removes Gaussian, stripe and deadline noise and preserves the detailed information of the hyperspectral images.