Facial Mask Detection on Live Camera Feed : Development of Robust Object Detection Algorithms for Manufacturing

Abstract: Modern Image Classification and Object Detection methods are ideal candidates for performingNon-Destructive Testing (NDT) in manufacturing establishment. These methods provide anopportunity to achieve digitalization and can form the backbone of Industry 4.0 implementationsin manufacturing environments. As with all applications in manufacturing productionenvironments, accuracy and detection speeds form important parameters for evaluation ofthese techniques.As part of this study, a set of generic models has been developed with the requirements ofmanufacturing environments in mind. These models vary from custom Convolutional NeuralNetwork (CNN) models to sophisticated modern Object Detection models like Single ShotDetection (SSD) and You Only Look Once Version 5 (YOLOv5). To conduct this study andshowcase the applicability of these models to a wide array of datasets, a custom datasetcreated for this study was used. This dataset had images of individuals wearing facial masksand not wearing masks. These models were tested on their ability to detect the presence of masks on a person´s face in a real time video feed. The evaluation on speed of detection was carried out on a variety of machines ranging from High end GPU enabled computers to portabledevices with compromised computing capabilities.The principal finding of this work is that several different models perform well with respect toaccuracy. However, most of these models suffer from the drawback that their speed of detectionon a live video is low, which in a production environment would lead to potential loss ofproductivity. The object detection model which performed best in terms of accuracy and speedin the current study was the YOLOv5 model. AB Sandvik Materials Technology, who providedthe support for this project, has also evaluated porting this YOLOv5 model to their applicationsin productions as part of a successful Proof of Concept study.