Improvement of a Dataset for the Creation of an Automatic Mask and Respirator Detection System

The COVID-19 pandemic made the whole world change its lifestyle in several respects: working at home has become much more common and people have started to pay more attention to safety. Social distancing and the use of protective masks helped the population to contain the spread of the epidemic, proving to be an effective means of dealing with this type of problem. Enforcement has been entrusted to staff dedicated to this task, but it is a job easily replaced through the use of automation. The FMR-DB (Facial Masks and Respirators Database) was developed in order to provide a basis for the training of machine learning systems capable of performing this task not only in the context of disease containment but in all contexts in which the use of masks plays a key role in safety protection. It consists of 4200 images collected using various search engines and the images are divided according to the presence or absence of masks or respirators and occlusions within the image. The images were selected to offer diversity in the selection of the subjects depicted, trying to represent different sexes, ages, and ethnicities to offer as universal a model as possible. The labeling process was carried out by means of LabelImg, a graphical image annotation tool, and the label files are saved in PascalVOC and YOLOv7 format: two of the most common labeling systems were chosen to give the possibility of adapting the database to different possible uses according to the desired architecture. Although both formats were chosen, the FMR-DB was built with the aim of training a model using YOLOv7, a deep learning-based recognition system based on the YOLO family of models that have become widely used in the industry in the last few years. YOLOv7 has been chosen because both its speed and accuracy are state-of-the-art and because it offers a range of models designed to be used on different types of hardware. A recognition system trained with FMR-DB could be used to ensure safety in the working environment, in hospitals, and in all contexts where the use of respirators prevents damage to health. The database could be improved by expanding the image collection to cover different lighting conditions and perspectives to cover extreme cases.