Efficient and accurate face mask detection with active learning and optimized convolutional neural networks with transfer learning

The global COVID-19 outbreak made our society aware of the significant role that mask-wearing plays in the prevention of viral transmission. Almost everywhere, world health authorities have been recommending the use of face masks in public spaces, with some even making it mandatory. Therefore, a significant rush is underway to develop automated face mask detection systems for surveillance purposes in areas such as transportation systems, shopping malls, and educational institutions, with the aim of monitoring the implementation of face mask policies. This work introduces a unique approach to enhance face mask detection by combining an Active Learning (AL) system with a Convolutional Neural Network (CNN), and fine-tuning the CNN's hyperparameters using a Genetic Algorithm (GA). We use the AL framework to query the most informative data samples, which not only minimizes the labelling cost but also achieves high model accuracy. To improve CNN's performance, hyperparameter optimization uses a genetic algorithm to optimally select the network parameters. The study leverages transfer learning and pruning on the CNN model to improve results. Pruning simplifies the network for faster inference, while transfer learning increases accuracy by leveraging the weights and biases of previously trained models. Benchmark datasets assess the proposed method, demonstrating its superior performance in face mask detection with higher accuracy and robustness compared to previous methods. According to the experiment, there are different levels of accuracy in training different active learning sampling strategies that use the transfer of learnt CNN pruned. The entropy sampling method outperforms all other methods, achieving an accuracy of 98%. We compared the transfer-learned pruned CNN model with the Corona mask two-stage CNN model and the fine-tuned Yolov6 model for real-time face mask recognition.