A Novel Remote Sensing Image Registration Algorithm Based on Feature Using ProbNet-RANSAC.

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Feature-based image registration is robust but susceptible to noise and occlusions affecting feature point accuracy.
• Traditional Random Sample Consensus (RANSAC) methods can degrade in performance with contaminated feature point sets.

Purpose of the Study:

• To develop a semi-automated method for generating image registration training data to enable deep neural network training.
• To enhance the RANSAC algorithm for more accurate transformation model estimation by incorporating learned guidance for hypothesis sampling.

Main Methods:

• A semi-automated method for creating image registration training data was developed.
• A probabilistic formulation of RANSAC was presented, guiding hypothesis sampling.
• ProbNet, a deep convolutional neural network, was built to generate sampling probabilities for feature points, optimizing RANSAC's minimum set selection.

Main Results:

• Qualitative experiments demonstrated the method's effectiveness through checkerboard visualizations.
• Quantitative experiments showed superior performance compared to vanilla RANSAC, LMeds-RANSAC, and ProSAC-RANSAC across seven evaluation metrics.
• The integration of model estimation within a deep learning framework allowed for end-to-end optimization, further improving registration accuracy.

Conclusions:

• The proposed ProbNet-guided RANSAC method significantly improves image registration accuracy and robustness.
• The semi-automated data generation and deep learning integration offer a scalable and effective approach for training image registration models.
• This work advances feature-based image registration by enhancing the reliability of transformation model estimation in challenging conditions.