A chi-squared-transformed subspace of LBP histogram for visual recognition

Area of Science:

• Computer Vision
• Machine Learning
• Pattern Recognition

Background:

• Local Binary Pattern (LBP) and its variants are prevalent in recognition tasks.
• Subspace methods applied to LBP are optimal for Gaussian distributions, but LBP features are non-Gaussian.
• Existing methods struggle with the inherent non-Gaussian nature of LBP features.

Purpose of the Study:

• To develop a novel feature transformation to better adapt LBP features for subspace analysis.
• To enhance recognition accuracy in various tasks by improving feature representation.
• To address the limitations of traditional subspace methods when applied to non-Gaussian LBP features.

Main Methods:

• A chi-squared transformation (CST) was proposed to map LBP features closer to a Gaussian distribution.
• Asymmetric Principal Component Analysis (APCA) was applied to the transformed features to handle their asymmetric nature and remove unreliable dimensions.
• The combined CST-APCA method was evaluated on spatial LBP for face recognition, protein cellular classification, and spatial-temporal LBP for dynamic texture recognition.

Main Results:

• The proposed chi-squared transformation (CST) effectively transforms LBP features towards a Gaussian distribution.
• Asymmetric Principal Component Analysis (APCA) successfully addressed the asymmetric nature of the transformed features.
• Extensive experiments demonstrated significant improvements in recognition accuracy across all tested tasks.

Conclusions:

• The CST-APCA method offers a robust approach for enhancing LBP-based recognition systems.
• Feature transformation is crucial for optimizing subspace analysis of non-Gaussian features like LBP.
• The proposed method shows broad applicability and effectiveness in diverse recognition domains.