Face Recognition on Low Resolution CCTV Video using GhostFaceNets

Abstract

This research aims to develop a face recognition system for low-resolution CCTV cameras in school areas, as part of the Jagai Anakta' program initiated by the government of Makassar. The system is based on the GhostFaceNets algorithm, which uses ghost modules, feature fusion, and attention mechanism to achieve high accuracy and efficiency in face recognition. The system is trained and evaluated on a custom dataset collected from SD Telkom Makassar, which comprises 18 individuals, including 10 students and 8 teachers. The dataset is preprocessed and augmented to enhance its diversity and robustness. The system's performance is measured using three evaluation stages: loss and accuracy based on the ArcFaceLoss matrix, loss and accuracy using .bin files, and accuracy and suitability using CCTV video. The results show that the system can handle up to five objects in the real-time frame with high confidence, but it faces difficulties when the number of objects increases. This study advances the GhostFaceNets algorithm for enhanced performance in low-resolution image processing and multi-object detection in real-time scenarios. Traditional benchmarks of GhostFaceNets primarily involve single-face detection in still images. This research extends this by adapting the architecture to effectively handle multiple faces in real-time video feeds. Key modifications include the removal of the DFC attention branch to prevent significant data representation loss and adjustments in the number of embedding layers. The integration of L2 regularization and the use of PReLU activation functions further refine the model's training effectiveness on the custom-compiled dataset. The system also shows a high degree of generalization and adaptability to different environmental conditions and scenarios. The system can be a useful tool for monitoring and providing early warning of criminal activities in school areas using a data-driven approach.