False Discovery Rate Approach to Unsupervised Image Change Detection.

Area of Science:

• Computer Vision
• Image Analysis
• Statistical Modeling

Background:

• Change detection in coregistered images is crucial for monitoring.
• Existing methods often require supervision or struggle with large-scale testing.
• Unsupervised approaches are needed for efficiency and broad applicability.

Purpose of the Study:

• To develop an unsupervised change detection method for multiple coregistered images.
• To employ an empirical-Bayesian approach with a false discovery rate (FDR) formulation.
• To enable efficient statistical inference for change detection in large-scale image analysis.

Main Methods:

• Utilized an empirical-Bayesian framework for statistical inference on local image patches.
• Implemented a false discovery rate (FDR) control for robust hypothesis testing.
• Applied rank-based statistics (Wilcoxon, Cramér-von Mises, Levene) for feature extraction.
• Designed an unsupervised change detector assuming limited changes in imagery.

Main Results:

• Demonstrated accurate performance across diverse datasets, including radar, dermatological, and surveillance imagery.
• Showcased the flexibility of the method in addressing application-specific detection problems.
• Validated the effectiveness of the empirical-Bayesian approach and FDR control.

Conclusions:

• The proposed unsupervised change detection method is accurate and flexible.
• The empirical-Bayesian approach with FDR offers an efficient solution for large-scale image analysis.
• The method shows promise for various real-world applications requiring change monitoring.