Angle 2DPCA: A New Formulation for 2DPCA

Area of Science:

• Computer Vision
• Machine Learning
• Data Science

Background:

• Principal Component Analysis (PCA) and its 2-Dimensional variant (2DPCA) are standard dimensionality reduction techniques.
• Traditional 2DPCA utilizes the L2-norm (squared Euclidean distance), making it sensitive to outliers in data representation and classification.
• Outliers can significantly degrade the performance of dimensionality reduction algorithms, impacting downstream tasks like image recognition.

Purpose of the Study:

• To address the sensitivity of 2DPCA to outliers by proposing a novel formulation called Angle-2DPCA.
• To develop a robust dimensionality reduction method that enhances data representation and classification accuracy.
• To introduce an efficient algorithm for solving the Angle-2DPCA optimization problem.

Main Methods:

• Developed Angle-2DPCA, a new formulation of 2DPCA that employs the L1-norm as the distance metric.
• Incorporated the relationship between reconstruction error and variance into the objective function for improved robustness.
• Designed a fast iterative algorithm to efficiently compute the solutions for Angle-2DPCA.

Main Results:

• Angle-2DPCA demonstrated superior performance compared to traditional 2DPCA in handling outliers.
• Experimental results on benchmark face image databases (Extended Yale B, AR, PIE) validated the effectiveness of the proposed approach.
• The L1-norm based metric in Angle-2DPCA provided more robust feature extraction.

Conclusions:

• Angle-2DPCA offers a robust alternative to standard 2DPCA for dimensionality reduction, particularly in the presence of outliers.
• The proposed method effectively improves data representation and classification accuracy in image analysis tasks.
• The developed iterative algorithm provides an efficient means to implement Angle-2DPCA for practical applications.