Hyperspectral Image Denoising via Matrix Factorization and Deep Prior Regularization.

Area of Science:

• Computer Vision
• Signal Processing
• Machine Learning

Background:

• Deep learning for 2D image denoising is established, but hyperspectral image (HSI) denoising faces challenges.
• High computational complexity and the need for universal 3D training datasets hinder end-to-end deep learning for HSI denoising.

Purpose of the Study:

• To develop an efficient hyperspectral image denoising framework for mixed Gaussian impulse noise removal.
• To overcome the limitations of end-to-end deep learning by proposing a CNN-constrained non-negative matrix factorization (NMF) approach.

Main Methods:

• Modeled HSI denoising as a CNN-constrained NMF problem.
• Utilized proximal alternating linearized minimization for optimization, involving spectral matrix update, abundance matrix update, and sparse noise estimation.
• Designed a CNN architecture with two training schemes enabling training on 2D image datasets.

Main Results:

• The proposed method demonstrates effective removal of Gaussian and mixed Gaussian impulse noises.
• Achieved comparable or superior performance against state-of-the-art denoising methods.
• The trained CNN model can denoise HSIs with varying channel numbers after a single training session on 2D data.

Conclusions:

• The CNN-constrained NMF framework offers a practical and efficient solution for HSI denoising.
• The ability to train on 2D datasets significantly reduces computational complexity and data requirements.
• This approach provides a versatile tool for HSI noise reduction across different spectral resolutions.