Blind Deblurring of Remote-Sensing Single Images Based on Feature Alignment

Area of Science:

• Remote Sensing Image Processing
• Computer Vision
• Deep Learning

Background:

• Motion blur significantly degrades remote sensing image quality, impacting detection and recognition accuracy.
• Existing deep learning methods for motion deblurring often struggle with feature misalignment, leading to lost details and errors.

Purpose of the Study:

• To propose an end-to-end generative adversarial network (SDD-GAN) for optimizing single-image motion deblurring in remote sensing.
• To address feature misalignment and improve detail recovery in blurred remote sensing images.

Main Methods:

• Developed SDD-GAN incorporating a feature alignment module (FAFM) to correct feature map offsets.
• Introduced a feature importance selection module to preserve reliable spatial and channel domain details.
• Constructed a novel remote sensing dataset (RSDATA) simulating satellite motion blur.

Main Results:

• SDD-GAN demonstrated superior performance over existing algorithms on both synthetic and real-world remote sensing datasets.
• The proposed feature alignment and importance selection modules effectively mitigated feature misalignment and enhanced detail preservation.
• Experiments on real satellite imagery (CX-6(02)) confirmed the algorithm's practical effectiveness.

Conclusions:

• SDD-GAN offers a robust solution for motion blur recovery in remote sensing imagery.
• The method significantly improves image quality, leading to better detection and recognition capabilities.
• The developed techniques provide a valuable advancement for processing high-speed orbital motion-blurred images.