Detecting Matching Blunders of Multi-Source Remote Sensing Images via Graph Theory

Area of Science:

• Computer Vision
• Photogrammetry
• Remote Sensing

Background:

• Multi-source image matching faces challenges due to radiometric and geometric distortions, leading to fewer accurate matching points and high blunder ratios.
• Traditional blunder detection methods struggle with inexplicit global geometric relationships and limited matching data.

Purpose of the Study:

• To develop robust matching blunder detection methods for multi-source images, addressing limitations of existing techniques.
• To improve the accuracy and reliability of image matching in challenging scenarios with significant distortions.

Main Methods:

• Proposed two graph theory-based methods: a complete graph-based approach using matched triangles and a TIN graph-based approach for computational efficiency.
• These methods leverage local geometric similarity constraints to identify and remove erroneous matches.

Main Results:

• Both graph-based methods demonstrated superior performance over the Random Sample Consensus (RANSAC) method in recognition rate, false rate, and positional accuracy across simulated and real multi-source image data.
• The TIN graph-based method achieved a mean false rate of 0.14 and mean positional accuracy (RMSE) of 1.5 pixels on real data, significantly outperforming RANSAC (0.50 false rate, 2.6 RMSE).
• The TIN graph-based method offers comparable computation time to RANSAC for blunder ratios up to 50% while being substantially faster than the complete graph method.

Conclusions:

• Graph theory-based methods provide effective solutions for matching blunder detection in multi-source remote sensing imagery, especially under conditions of low matching points and high blunder ratios.
• The proposed complete graph and TIN graph methods offer significant improvements in accuracy and reliability for image matching applications.