Local Sliced Wasserstein Feature Sets for Illumination Invariant Face Recognition

Area of Science:

• Computer Vision
• Image Processing
• Pattern Recognition

Background:

• Face recognition systems often struggle with variations in illumination.
• Existing methods may not adequately address complex lighting deformations in facial images.

Purpose of the Study:

• To develop a robust face recognition method resilient to varying illumination conditions.
• To model and compensate for illumination-induced deformations in local image gradients.

Main Methods:

• Utilizing the Radon Cumulative Distribution Transform (R-CDT) for mathematical modeling of local gradient distributions.
• Representing illumination-induced deformations as a subspace in the R-CDT domain.
• Employing a nearest subspace method for face recognition in the R-CDT domain.

Main Results:

• The proposed R-CDT-based method demonstrates superior performance in face recognition tasks with challenging illumination.
• Experimental results confirm the effectiveness of modeling gradient distribution deformations as a subspace.
• The method significantly outperforms alternative approaches under varied lighting.

Conclusions:

• The R-CDT-based approach offers a robust solution for face recognition under difficult illumination.
• The mathematical modeling of gradient distributions provides a powerful framework for handling illumination variations.
• Publicly available Python code (PyTransKit) facilitates the implementation and adoption of this method.