Discriminant Analysis via Joint Euler Transform and ℓ2,1-norm.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Linear Discriminant Analysis (LDA) is a common technique for face recognition.
• Outliers in real-world data can significantly degrade LDA performance.
• Existing methods struggle with noisy datasets, limiting practical applications.

Purpose of the Study:

• To propose a robust distance metric learning method for Linear Discriminant Analysis (LDA) that effectively handles outliers.
• To enhance face recognition accuracy in unconstrained environments.
• To develop a computationally efficient algorithm for the proposed method.

Main Methods:

• Introduced Euler LDA-L21 (e-LDA-L21), a two-stage method involving Euler transform to a complex space and adopting the ℓ2,1-norm as a distance metric.
• Developed an iterative algorithm with guaranteed convergence and closed-form solutions for efficient computation.
• Extended the method to Euler 2DLDA-L21 (e-2DLDA-L21) to incorporate spatial image information.

Main Results:

• The proposed e-LDA-L21 method demonstrates superior performance compared to state-of-the-art algorithms.
• Experimental results on multiple face databases validate the effectiveness of the approach in handling outliers.
• The Euler transform and ℓ2,1-norm effectively reveal nonlinear features and exploit data geometry.

Conclusions:

• Euler LDA-L21 (e-LDA-L21) offers a robust and effective solution for face recognition in the presence of outliers.
• The method enhances accuracy and reliability, paving the way for real-world deployment.
• The extension to e-2DLDA-L21 further improves performance by leveraging spatial image data.