[Paper Review] Local Sliced-Wasserstein Feature Sets for Illumination-invariant Face Recognition

Yan Zhuang, Shiying Li+4 more|arXiv (Cornell University)|Feb 22, 2022

Face recognition and analysis4 citations

TL;DR

This paper proposes a novel illumination-invariant face recognition method using Local Sliced-Wasserstein Feature Sets based on the Radon Cumulative Distribution Transform (R-CDT). By modeling local image gradient distributions as 2D discrete measures in R-CDT space, the method captures illumination-induced deformations as convex subspaces, enabling robust recognition via nearest subspace classification. The approach achieves state-of-the-art performance on challenging illumination datasets, outperforming existing methods in multiple benchmarks.

ABSTRACT

We present a new method for face recognition from digital images acquired under varying illumination conditions. The method is based on mathematical modeling of local gradient distributions using the Radon Cumulative Distribution Transform (R-CDT). We demonstrate that lighting variations cause certain types of deformations of local image gradient distributions which, when expressed in R-CDT domain, can be modeled as a subspace. Face recognition is then performed using a nearest subspace in R-CDT domain of local gradient distributions. Experiment results demonstrate the proposed method outperforms other alternatives in several face recognition tasks with challenging illumination conditions. Python code implementing the proposed method is available, which is integrated as a part of the software package PyTransKit.

Motivation & Objective

• Address the persistent challenge of illumination variation in face recognition, which can degrade performance more than identity changes.
• Develop a theoretically grounded, transport-based feature representation that is invariant to illumination-induced deformations of local image gradients.
• Enable robust face recognition under extreme lighting conditions using a mathematically principled subspace learning framework.
• Provide a method that is computationally efficient, data-efficient, and generalizable across diverse illumination datasets.
• Integrate preprocessing techniques like log transformation and explore adaptive weighting of image patches for performance gains.

Proposed method

• Divide input face images into local patches and compute 2D discrete gradient distributions for each patch.
• Apply the Radon Cumulative Distribution Transform (R-CDT) to map local gradient distributions into a transport-friendly space with convexifying properties.
• Model illumination-induced deformations of gradient distributions as linear transformations (scaling and translation) in the R-CDT domain.
• Represent each image as a set of R-CDT-transformed gradient features and aggregate them using weighted distance summation.
• Perform classification via nearest subspace method in the R-CDT feature space, leveraging the convexity of illumination-variant distributions.
• Incorporate log-pixel transformation as a preprocessing step to enhance robustness to illumination changes.

Experimental results

Research questions

• RQ1 Can illumination-induced deformations of local image gradient distributions be effectively modeled as a convex subspace in a transport-based feature space?
• RQ2 Does the R-CDT transformation enable better separation of illumination-variant gradient distributions compared to standard feature descriptors?
• RQ3 To what extent does the proposed method outperform existing illumination-invariant face recognition techniques on benchmark datasets with extreme lighting variations?
• RQ4 How does the inclusion of log-transformed preprocessing improve the robustness of the R-CDT-based feature representation?
• RQ5 Can adaptive weighting of local image patches further enhance recognition accuracy by emphasizing more discriminative regions?

Key findings

• The proposed method achieves superior performance on three benchmark datasets with challenging illumination variations, outperforming existing state-of-the-art approaches.
• In the Yale Face database, the method achieved a classification accuracy of 82.51% when training data included images from test subsets 1 and 2, up from 49.7% with only clean images.
• The addition of log-pixel preprocessing consistently improved classification accuracy, with the largest gains observed in regions of high illumination variation, such as test subset 4 of the Yale dataset.
• Theoretical analysis confirms that local gradient distributions under illumination variations form a convex set in the R-CDT domain, enabling effective subspace classification.
• Performance improves with increasing training data, indicating the method captures complex lighting variations beyond simple scaling and translation.
• The method demonstrates strong generalization and data efficiency, with significant accuracy gains even when limited training data is available.