Two-Dimensional Quaternion Sparse Discriminant Analysis.

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Linear discriminant analysis (LDA) is widely used for dimension reduction and feature extraction.
• Existing LDA methods face challenges in effectively representing and processing multi-channel image data like RGB and RGB-D.

Purpose of the Study:

• To propose a novel method, two-dimensional quaternion sparse discriminant analysis (2D-QSDA), for enhanced feature extraction from RGB and RGB-D images.
• To improve the generalization ability and preserve spatial structure in image data.

Main Methods:

• Incorporation of sparse regularization to focus on important variables, enhancing out-of-sample data generalization.
• Utilization of quaternion representation to preserve high-order correlations among image channels, unifying RGB and RGB-D feature extraction.
• Matrix-based processing to retain the spatial structure of input images.
• Transformation of the constrained trace ratio problem into a quaternion sparse regression (QSR) model, solved via a nested iterative algorithm in complex space.
• Development of 2D-QSDAw, incorporating weighted pairwise between-class distances to further improve separability.

Main Results:

• 2D-QSDA demonstrates strong generalization ability on unseen data due to sparse regularization.
• Quaternion representation effectively captures inter-channel correlations, providing a unified feature extraction approach.
• Experimental results on RGB and RGB-D databases validate the effectiveness of 2D-QSDA and 2D-QSDAw over existing methods.

Conclusions:

• 2D-QSDA offers a robust framework for feature extraction from RGB and RGB-D images by leveraging sparse regularization and quaternion representation.
• The proposed 2D-QSDAw variant further enhances feature separability, showing superior performance in experimental evaluations.
• This work advances discriminant analysis techniques for complex image data analysis.