Regularizing Hyperspectral and Multispectral Image Fusion by CNN Denoiser.

Area of Science:

• Remote Sensing
• Image Processing
• Computer Vision

Background:

• Hyperspectral image (HSI) and multispectral image (MSI) fusion is crucial for obtaining high-resolution HSI (HR-HSI) from low-spatial-resolution HSI (LR-HSI) and higher-resolution MSI.
• Existing fusion methods face challenges in accuracy and efficiency.

Purpose of the Study:

• To present a novel HSI and MSI fusion method, CNN-Fus, that leverages subspace representation and a convolutional neural network (CNN) denoiser.
• To enable direct application of the method to various HSI and MSI datasets without retraining the CNN.

Main Methods:

• The proposed CNN-Fus method utilizes subspace representation to approximate HR-HSI, learning the subspace from LR-HSI via singular value decomposition.
• A pre-trained CNN denoiser, originally developed for grayscale images, is integrated into the Alternating Direction Method of Multipliers (ADMM) algorithm to regularize coefficient estimation.

Main Results:

• The CNN denoiser, trained on accessible grayscale images, is directly applied to HSI and MSI fusion tasks.
• The method demonstrates faster processing and more accurate recovery of HR-HSI compared to state-of-the-art fusion techniques.

Conclusions:

• CNN-Fus offers a powerful and efficient approach for HSI and MSI fusion.
• The integration of a general-purpose CNN denoiser simplifies the fusion process and enhances performance across different datasets.