A Comparative Study of Transfer Learning and Custom Models with SSD for Face Mask Detection

Monitoring crowded public places, workplaces, and transportation hubs for compliance with safety guidelines and occupational health standards presents a significant challenge. With advancements in Deep Learning (DL) and video surveillance, automatic detection and classification of face masks have made substantial progress in computer vision and image processing. In this paper, we investigate and compare two architectures for detecting face masks: a custom neural network (NN) and a transfer learning approach using a pre-trained Inception model. Both architectures consist of two stages. First, faces are detected using a pre-trained Single Shot MultiBox Detector (SSD). Then, these faces are classified as masked or unmasked using a sequential model. We conducted various experiments to evaluate the performance and effectiveness of each approach using the public Face Mask Detection Dataset. Both techniques achieved high accuracy, exceeding 90 %, with the custom model surpassing 95%. With additional resources and training on larger datasets, the results could be further improved.