Detecting Masked Faces using Region-based Convolutional Neural Network

Precisely detecting masked and non-masked faces are increasingly important since wearing a face mask is an effective measure to prevent the spread of the COVID-19 pandemic. Previous literature focused mainly on occluded face detection or facial expression, which were unsuitable for the application of detecting human faces wearing masks. Hence, to overcome the issue, this work proposes a Convolutional Neural Network-based model that uses region proposals to detect masked and nonmasked faces. The depth of the Convolutional Neural Network was increased by using residual skip-connections. The model was implemented using TensorFlow Object Detection API and was pre-trained over the COCO dataset. A secondary outcome of the paper was to collect a dataset of high resolution masked and nonmasked faces for training deep learning frameworks, due to the lack of datasets available for this task. The proposed model was compared with the SSD Inception V2 model [1] and the SSD MobileNet V2 model [2] in the context of detection accuracy (mAP). The experiments highlight that the proposed framework achieves a detection accuracy (total mAP) of 85.82%, on the collected dataset. The results are significantly better in detecting non-masked faces with a detection accuracy (mAP) of 98.61% while it is 68.72% accurate in detecting masked faces. Based on this model’s performance on standard parameters, a detailed study is outlined along with the conclusion and future plan of action.