Bayesian Pairwise Comparison of High-Dimensional Images.

Area of Science:

• Statistical image analysis
• Computational statistics
• Machine learning

Background:

• Automated image analysis requires robust image pair comparison techniques.
• High-dimensional images present challenges due to pixel-wise differences, random error, and spatial dependencies.

Purpose of the Study:

• To develop an effective statistical method for automatically labeling image pairs as similar or different.
• To address the challenges of comparing high-dimensional images by accounting for spatial dependencies and random variability.

Main Methods:

• Developed a novel Bayesian strategy using an extension of Dirichlet processes called the spatial random partition model (sRPM).
• The sRPM groups spatially proximal pixels with similar intensities, achieving dimension reduction.
• Formulated image comparison as a hypothesis test using a univariate metric robust to spatial correlations and pixel intensity variability.
• Employed a two-stage MCMC (Markov chain Monte Carlo) analysis for computational efficiency.

Main Results:

• Simulation studies on artificial datasets demonstrated the high accuracy of the sRPM in image comparison.
• The sRPM effectively handles spatial correlations and random variability in pixel intensities.

Conclusions:

• The spatial random partition model (sRPM) provides an effective and accurate method for comparing image pairs.
• The technique is validated through statistical analysis of satellite image data, showing practical applicability.