Robust registration of SAR and optical images based on deep learning and improved Harris algorithm

Area of Science:

• Remote Sensing
• Computer Vision
• Machine Learning

Background:

• Traditional image registration algorithms struggle with multimodal images like SAR and optical data due to differing textures and structures.
• Accurate registration is crucial for fusing information from different imaging modalities.

Purpose of the Study:

• To develop a robust framework for accurate registration between synthetic aperture radar (SAR) and optical images.
• To overcome the limitations of existing methods in handling significant differences in image characteristics.

Main Methods:

• A novel deep learning network generates high-quality pseudo-optical images from SAR images.
• Multi-scale Harris algorithm detects and extracts feature points.
• Gradient Position Orientation Histogram (GPOH) constructs feature descriptors.
• A feedback mechanism reconstructs feature point positions for precise matching.

Main Results:

• The proposed framework demonstrates superior matching performance compared to state-of-the-art methods.
• The deep learning approach effectively bridges the modality gap between SAR and optical images.
• Accurate feature point detection and reconstruction are key to successful registration.

Conclusions:

• The presented robust matching framework significantly improves SAR and optical image registration.
• The integration of deep learning for pseudo-image generation is effective in addressing multimodal registration challenges.
• This work offers a promising solution for applications requiring accurate fusion of SAR and optical imagery.