Image Outlier Detection and Feature Extraction via L1-Norm-Based 2D Probabilistic PCA

Area of Science:

• Machine Learning
• Computer Vision
• Statistical Modeling

Background:

• Principal Component Analysis (PCA) is a widely used dimensionality reduction technique.
• Traditional PCA methods often assume Gaussian noise, limiting their performance with non-Gaussian data distributions.
• Outliers in datasets can significantly degrade the performance of PCA-based models.

Purpose of the Study:

• To introduce a novel L1-norm-based probabilistic principal component analysis model for 2D data, termed L1-2DPPCA.
• To leverage the Laplacian noise model for robust data analysis and outlier detection.
• To enhance image reconstruction accuracy and feature extraction performance compared to existing PCA methods.

Main Methods:

• Developed an L1-norm-based probabilistic PCA model (L1-2DPPCA) assuming a Laplacian noise model.
• Utilized the property that Laplacian density can be represented as a superposition of Gaussian distributions.
• Employed a Bayesian inference framework with a variational expectation-maximization algorithm for parameter learning.
• Introduced hidden variables within the Gaussian superposition to detect data outliers.

Main Results:

• The proposed L1-2DPPCA model effectively identifies data outliers using newly introduced hidden variables.
• Experimental results on public databases demonstrate significant performance improvements in image reconstruction after outlier removal.
• L1-2DPPCA generally outperforms existing PCA-based algorithms in feature extraction, showing lower reconstruction errors and higher classification accuracy.

Conclusions:

• L1-2DPPCA offers a robust approach for 2D data analysis, particularly when dealing with Laplacian noise.
• The model's ability to detect and handle outliers leads to superior image reconstruction and feature extraction.
• This method provides a valuable advancement for applications requiring accurate dimensionality reduction and pattern recognition in image data.