GRiD: Guided Refinement for Detector-Free Multimodal Image Matching.

Area of Science:

• Computer Vision
• Remote Sensing
• Geospatial Analysis

Background:

• Multimodal image matching is critical for applications like image stitching and land cover mapping.
• Nonlinear radiometric distortion (NRD) and geometric distortions significantly challenge existing multimodal image matching methods.
• Detector-based approaches suffer from feature point offset issues, especially in areas with large modal differences.

Purpose of the Study:

• To propose a novel guided refinement for detector-free multimodal image matching (GRiD) method.
• To address feature point offset issues and improve matching accuracy under severe distortions.
• To enhance the robustness and reliability of multimodal image matching.

Main Methods:

• Introduced a detector-free framework to establish pixel-level correspondences, bypassing feature point detection.
• Developed a guided correction module using robust reference points (RPs) to guide matching in challenging regions.
• Incorporated a phase congruency module to enhance RP reliability by concentrating them around edge structures.
• Performed finer localization on highly correlated corresponding pixels for optimized matches.

Main Results:

• The proposed GRiD method effectively weakens feature point offset issues.
• GRiD establishes robust pixel-level correspondences even with severe nonlinear radiometric distortion and geometric distortions.
• Extensive experiments on four multimodal datasets validated the method's sufficiency and robustness across various image modalities.

Conclusions:

• The GRiD method offers a significant advancement in multimodal image matching, particularly for images with substantial distortions.
• The approach effectively suppresses feature point offset problems, leading to more accurate registration and fusion.
• GRiD demonstrates strong potential for improving performance in diverse remote sensing and computer vision applications.