ULMR: An Unsupervised Learning Framework for Mismatch Removal

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Image matching is crucial but prone to mismatches due to distortions.
• Supervised deep learning methods require time-consuming manual labeling and are sensitive to labeling errors.

Purpose of the Study:

• To develop an unsupervised deep reinforcement learning framework for robust mismatch removal.
• To improve image matching accuracy and reduce reliance on labeled data.

Main Methods:

• Proposed an unsupervised learning for mismatch removal (ULMR) framework leveraging deep reinforcement learning (DRL).
• ULMR scores state-action pairs using a classification network, calculates policy gradients, and maximizes expected rewards to train the network.

Main Results:

• ULMR achieved higher precision, more correct matches, and fewer false matches compared to supervised and handcrafted methods.
• Demonstrated superior stability, accuracy, and quality in application experiments.

Conclusions:

• ULMR offers an effective unsupervised approach for image mismatch removal.
• The framework shows potential for reduced sampling times and enhanced compatibility with classification networks.