LR-Net: Low-Rank Spatial-Spectral Network for Hyperspectral Image Denoising

Area of Science:

• Remote Sensing
• Image Processing
• Computer Vision

Background:

• Hyperspectral images (HSIs) suffer from hybrid noise (Gaussian, impulse, stripes, dead lines) due to imaging device limitations.
• This noise significantly degrades the performance of crucial HSI applications like unmixing and classification.

Purpose of the Study:

• To propose a novel end-to-end deep learning network, LR-Net, for effective hybrid noise removal in HSIs.
• To leverage the inherent low-rank property of clean HSIs within a deep convolutional neural network (DCNN) framework.

Main Methods:

• Developed a novel end-to-end low-rank spatial-spectral network (LR-Net) integrating DCNNs with low-rank priors.
• Incorporated spatial-spectral atrous blocks (SSABs) and a multi-atrous block (MAB) to extract multi-level spatial-spectral features.
• Utilized a plug-and-play low-rank module (LRM) for differentiable low-rank matrix reconstruction and feature recovery.

Main Results:

• LR-Net demonstrated superior performance over state-of-the-art denoising methods on both simulated and real-world HSI datasets.
• Achieved significant improvements in quantitative evaluation metrics and visual quality assessments.
• Showcased strong stability and generalization capabilities in hybrid noise removal.

Conclusions:

• The proposed LR-Net effectively removes hybrid noise from HSIs by synergistically combining DCNNs and low-rank properties.
• This approach enhances the quality of HSIs, paving the way for improved downstream analysis and applications.
• LR-Net offers a robust and generalizable solution for HSI denoising challenges.