PRF-Net: A Progressive Remote Sensing Image Registration and Fusion Network.

Area of Science:

• Remote Sensing
• Computer Vision
• Image Processing

Background:

• Existing fusion algorithms struggle with unregistered or locally misaligned images, degrading fused image quality.
• Nonlinear misregistration persists even after standard image registration techniques.

Purpose of the Study:

• To propose a progressive remote sensing image registration and fusion network (PRF-Net) robust to image misalignment.
• To improve the quality of fused remote sensing images, especially for images from different platforms.

Main Methods:

• A registration network comprising a global spatial transform network (GSTN) for coarse alignment and a local spatial warp network (LSWN) for fine-tuning.
• A fusion network incorporating a multiscale feature extraction (MSFE) block and a spatial details attention (SDA) block to preserve spectral and spatial information.

Main Results:

• PRF-Net demonstrated excellent performance in both reduced and full resolutions across four types of remote sensing images.
• The network effectively handles local misregistration, leading to superior registration and fusion quality compared to existing methods.

Conclusions:

• The proposed PRF-Net effectively addresses the challenges of image registration and fusion for remote sensing data.
• The network's design preserves crucial spatial and spectral details, resulting in high-quality fused images.