Nonparametric Bayesian Correlated Group Regression With Applications to Image Classification.

Area of Science:

• Computer Vision
• Statistical Modeling
• Machine Learning

Background:

• Sparse Bayesian learning is widely used for image classification.
• Existing models often assume independent Gaussian noise, which is unrealistic.
• Practical image noise exhibits long-tail distributions and spatial correlations.

Purpose of the Study:

• To develop a novel sparse Bayesian learning method that accurately models complex image noise.
• To improve image classification performance by incorporating realistic noise characteristics.

Main Methods:

• Proposed Bayesian correlated group regression (BCGR) by partitioning noise into groups.
• Modeled each group using a long-tail distribution (scale mixture of matrix Gaussian).
• Employed nonparametric Bayesian estimation with low-rank and matrix Gamma priors on covariance matrices.

Main Results:

• BCGR effectively captures intragroup noise correlation and long-tail properties.
• The method integrates noise distribution, long-tail attributes, and structure information.
• Learned covariance matrices were used to construct a grouped Mahalanobis distance for classification.

Conclusions:

• The proposed BCGR method enhances data reconstruction by better fitting practical noise.
• BCGR demonstrates significant effectiveness in image classification tasks.
• The model automatically adapts to noise distributions and integrates structural information.