A Novel Approach to Face Recognition in Videos Based on a Single Reference Image

 

 

 

Mohammed Ahmed Talab1, Mustafa A. Feath2, Ahmed Hadi Ali AL-Jumaili3,*, Mohammed A. Al-shibl4 ,
Ravie Chandren Muniyandi5

 

1Department of medical Physics, College of Applied Science, University of Fallujah, Anbar, 00964, Iraq

 

2Director of the Department of Studies and Planning, College of Medicine, University of Anbar, Anbar, 00964, Iraq

 

3College of Information Technology, University of Fallujah, Anbar, 00964, Iraq

 

4Director of Computer Center, University of Fallujah Anbar, 00964, Iraq

 

5College of Computing & Informatics (CCI), Universiti Tenaga Nasional (UNITEN), Putrajaya Campus, Jalan IKRAM-UNITEN, 43000 Kajang, Selangor, Malaysia

 

Emails: mmss_ah@uofallujah.edu.iq; Azeezmustafa89@uoanbar.edu.iq; ahmed_hadi@uofallujah.edu.iq; dr.alshibly@uofallujah.edu.iq; ravie.chandren@uniten.edu.my

 

Text Box: Abstract This paper introduces an advanced method for face recognition in video surveillance systems, leveraging only a single reference image per individual. The challenge of recognizing faces in video is addressed, considering issues like pose variations, occlusions, and lighting changes. The proposed approach utilizes 3D Morphable Models (3DMM) to generate a 3D face mesh from the reference image, facilitating robust face alignment and recognition across video frames. A Convolutional Neural Network based pipeline is employed for face detection, pose estimation, and extraction of invariant features, while an optimization framework refines landmark positions and depth maps for accurate 3D reconstruction. The system performs exceptionally well on the CASIA-WebFace Dataset, with 97.00% pAUC (20%) in surveillance mode and 98.69% in identification mode for frontal views. With an efficiency of 16.72 FPS on modest hardware, the system proves its practicality for real-world deployment. The method incorporates synthetic data augmentation and Random Subspace Methods to enhance adaptability to domain-specific conditions. Compared to existing methods like Eoe-SVM and CCM-CNN, the proposed system demonstrates a superior balance between accuracy and computational efficiency, particularly in Single Sample Per Person (SSPP) scenarios. By focusing on single-reference image recognition, the system offers a promising solution for large-scale surveillance applications, where video footage typically contains multiple poses, expressions, and lighting variations. The results highlight the system's effectiveness and efficiency, making it an excellent alternative for real-time face recognition in complex and dynamic surveillance environments.

 

Received: January 19, 2025 Revised: March 06, 2025 Accepted: April 08, 2025

 

Keywords: Face recognition; video surveillance; 3D morphable models; single reference image; domain adaptation; Deep learning