CDFAN: Cross-Domain Fusion Attention Network for Pansharpening.

Area of Science:

• Remote Sensing
• Image Processing
• Computer Vision

Background:

• Pansharpening computationally enhances spatial resolution of low-resolution hyperspectral images (LRMS) using high-resolution panchromatic (PAN) data.
• Existing methods struggle with preserving high-frequency details and integrating spatial-spectral information effectively.

Purpose of the Study:

• To introduce a novel architecture, the Cross-Domain Fusion Attention Network (CDFAN), for improved pansharpening.
• To address limitations of traditional spatial-domain and frequency-based pansharpening approaches.

Main Methods:

• CDFAN employs a Multi-Domain Interactive Attention (MDIA) module using discrete wavelet transform (DWT) for cross-domain attention.
• A Spatial Multi-Scale Enhancement (SMCE) module with multi-scale convolutional pathways and an Expert Feature Compensator is utilized.
• Attention mechanisms are constructed across wavelet and spatial domains for effective fusion.

Main Results:

• CDFAN demonstrates significant improvements over state-of-the-art pansharpening methods on benchmark datasets.
• The proposed network achieves enhanced spectral-spatial fidelity in the fused high-resolution multispectral (HRMS) images.
• Results show superior reconstruction quality and perceptual quality.

Conclusions:

• CDFAN offers a robust solution for pansharpening by effectively integrating multi-domain information.
• The architecture successfully preserves high-frequency textures and spectral integrity.
• This advancement contributes to higher quality hyperspectral image fusion for various applications.