Definition of Unique Objects by Convolutional Neural Networks using Transfer Learning

Abstract

This article solves the problem of detecting medical masks on a person's face. Medical mask is one of the most effective measures to prevent infection with COVID-19, and its automatic detection is an actual task. The introduction of automatic recognition of medical masks in existing information security systems will allow quickly identify the violator of the mask regime, which in turn will increase security in a pandemic. The article provides a detailed analysis of existing solutions for face detection and automatic recognition of medical masks, method based on the use of convolutional neural networks was proposed. A distinctive feature of the new method is the use of two neural networks at once, using the RetinaFace neural network architecture at the face search stage and using the Resnet neural network architecture at the face mask recognition stage. It is shown that the use of transfer learning on scales, learned to work with faces, significantly accelerates learning and increases the accuracy of recognition. However, with this approach, there are some false positives, for example, when you try to cover your face with your hands, imitating a medical mask. Based on the study, we can conclude that the algorithm is applicable in the security system to determine the presence/absence of a medical mask on a person's face, as well as the need for additional research to solve the problems of false positives of the algorithm.