Abstract

This paper addresses the critical challenge of robust face recognition in biometric authentication systems under random partial occlusions, such as masks, glasses, scarves, or other obstructions. The primary contribution is the development of a novel Reference Sample Adaptation (RSA)-based normalization technique designed to improve recognition performance when facial images contain significant occlusions. The main objective is to develop a real-time system capable of accurately identifying users even when substantial portions of the face are occluded. The study analyzes limitations of both classical face recognition algorithms and modern AI-based methods, particularly their reduced robustness to partial visibility. To overcome these limitations and fill the existing research gap in occlusion-robust normalization, we propose an RSA-based normalization method.This approach first classifies face images, generates normalized reference images from clean samples of each individual, and then adaptively normalizes occluded input images before the recognition stage [3]. The proposed pipeline integrates triangle-based feature detection, pixel-level projections, and Viola–Jones features for robust landmark localization and obstruction removal, followed by deep learning models to extract highly discriminative features. This hybrid methodology advances beyond many state-of-the-art occlusion-handling techniques by combining classical preprocessing with deep feature extraction, while explicitly addressing the shortcomings of end-to-end deep detectors in highly occluded scenarios. Experimental results on a collected dataset demonstrate that the proposed RSA-based normalization significantly increases recognition accuracy and reduces False Acceptance Rate (FAR) and False Rejection Rate (FRR) compared to baseline methods. Further validation against contemporary benchmarks, including margin-based losses such as ArcFace and recent masked/occlusion-robust approaches published after 2020, confirms the effectiveness of the method. The developed system shows strong potential for real-time biometric authentication in secure access control, surveillance, and other environments where occlusions are common

Statistics