Cascaded YOLO v8 and ResNet-50 based Masked Face Detection

Abstract

Large population of the world have been affected due to respiratory infectious diseases. General recommendations for proper face masks usage, regular and proper hand hygiene as well as employing social distancing at such infection prone areas could help in managing the transmission of these illnesses. Masked face detection would be used as an essential tool for the monitoring as well as employment of different control measures. The works included in the article consists of two parts: bounding box estimation of the faces (with and without occlusions) that was accomplished by the implementation of modified YOLO v8. The second part was the implementation of ResNet-50 for classification of the faces. The models used was trained with a masked face dataset that included 853 images. The dataset consists of properly masked faces, improperly masked faces as well as unmasked faces. The modified YOLO v8 model with FPN was used for training of the available training set for bounding box estimation of the faces. Here, the neck of YOLO v8 (PANet) was replaced with the FPN. The mean average precision of the modified model was found to be better than that of the unmodified YOLO model. The mAP50 for the modified YOLO v8 was 0.877 which was better as compared to the unmodified YOLO v8 model with 0.847. Similarly, the use of ResNet-50 was done for classificaiton of the faces obtained from the bounding box estimated using YOLO models. The least training loss of the ResNet-50 was obtained to be 0.4744.