New robust face recognition methods based on linear regression.

Area of Science:

• Computer Vision
• Machine Learning
• Pattern Recognition

Background:

• Nearest Subspace (NS) classification, particularly linear regression-based classification (LRC), is an efficient face recognition technique.
• Existing NS methods struggle with the small sample size (SSS) problem, leading to misclassifications due to limited training data per class.
• The assumption that class samples lie within a specific subspace is fundamental but challenged by SSS.

Purpose of the Study:

• To propose two novel robust linear regression-based classification (RLRC) methods for enhanced face recognition.
• To address the limitations of existing LRC methods caused by the small sample size (SSS) problem.
• To develop a method that leverages the existence of class-specific basis vectors without explicit calculation.

Main Methods:

• The proposed methods, robust LRC 1 and 2 (RLRC 1 and 2), are based on the concept that class-specific subspaces are spanned by common and unique basis vectors.
• Unlike prior methods, RLRC 1 and 2 do not require the explicit extraction of class-specific basis vectors.
• These methods utilize downsampled face images as features for classification.

Main Results:

• Experiments on three benchmark face databases demonstrate the superior performance of RLRC 1 and 2.
• The proposed methods show significant improvements in face recognition accuracy compared to existing state-of-the-art techniques.
• RLRC 1 and 2 provide more robust face recognition, effectively mitigating the SSS problem.

Conclusions:

• The novel RLRC 1 and 2 methods offer a more robust and accurate solution for face recognition, particularly under conditions of limited training data.
• These methods advance the field of subspace-based classification by effectively handling the SSS challenge.
• The proposed approach validates the effectiveness of utilizing the concept of basis vectors without direct computation for improved recognition performance.