Boosting Time Efficiency in Helmeted Face Recognition with Transfer Learning and k-NN

Abstract

This research aims to optimize the retraining process of face recognition systems for motor vehicle parking areas using Transfer Learning and k-nearest Neighbors (k-NN). The system enhances efficiency by retraining only on new classes, identified by checking the encoding file, allowing it to skip pre-existing classes, thus significantly improving retraining speed. Face data from 100 individuals (2725 images) was collected and augmented, with the VGG16 architecture used to generate face embeddings through Transfer Learning. Achieving 95.62% accuracy, with a precision, recall, and F1-score of 0.96 at k=1, the system demonstrates high accuracy and speed in identifying registered faces while effectively rejecting unknown ones. These optimizations greatly increase the feasibility of face recognition in parking security systems, making them more responsive to new data. Future research should focus on increasing dataset diversity and exploring more complex conditions, such as varied lighting, to enhance system performance and robustness further.