FaceLite: A Real-Time Light-Weight Facemask Detection Using Deep Learning: A Comprehensive Analysis, Opportunities, and Challenges for Edge Computing

The edge computing devices running models based on deep learning have drawn a lot of interest as a prominent way of handling various applications based on AI. Due to limited memory and computing resources, it is still difficult to deploy deep learning models on edge devices in a production context with effective inference. This study examines the deployment of a lightweight facemask detection model on edge devices with real-time inference. The proposed framework uses a dual-stage convolutional neural network (CNN) architecture with two main modules that use Caffe-DNN for face detection and a proposed model based on CNN architecture or customized models based on transfer learning (e.g., MobileNet-v2, resNet50, denseNet121, NASNetMobile, Inception-v3, and XceptionNet) for facemask classification. The study does numerous analyses based on the models' performance in terms of accuracy, precision, recall, and F1-score and compares all models with low disk size and good accuracy as the main priorities for memory-constrained edge devices. The proposed CNN model for facemask detection outperforms other state-of-the-art models in terms of accuracy, achieving 99%, 99%, and 99% on the training, validation, and testing, respectively, with the facemask detection ~12K image datasets available on Kaggle. This accuracy is comparable to other transfer learning-based models, and it also achieves the smallest number of total trainable parameters and, thus, the smallest disk size of all models.