Using the Yolov5 Neural Network to Recognize the Presence of Personal Protective Equipment

Abstract

The article discusses the application of the YOLO neural network for the automatic detection and identification of personal protective equipment for enterprise personnel in digital images. A comparative analysis of various YOLOv5 network architectures is given, depending on the neural network learning rate and recognition accuracy. The tools for training a neural network in Python within the Google Colab environment, which allows efficient use of remote computing power of GPUs, are considered. An example of network training based on collected dataset with six classes for recognizing the presence of personal protective equipment (with a mask; without a mask; the mask is worn incorrectly; with a head protector; without a head protector; special clothing) is considered. The settings of the YOLOv5 neural network and recommendations for improving object recognition by supplementing the dataset with certain condition images are described. Four variants of network training and different number of epochs were produced using the YOLOv5s and YOLOv5l architecture. The optimal neural network architecture and the number of required epochs has been selected. The concept of average accuracy (Mean Average Precision - mAP) is considered and a comparative analysis of the prediction accuracy from the selected network architecture and settings is given. A practical example of recognizing the presence of personal protective equipment in digital images using a trained YOLOv5 neural network is given. On average, the recognition accuracy is 0.7, which is not a bad result for solving this class of problems. The paper deals with issues of improving the accuracy of object recognition. Plans for the detection of previously detected and new classes of personal protective equipment are given, as well as confirmation of the fact that the personnel of enterprises have passed the regulatory procedures based on digital video sequence.