Using Convolutional Neural Networks to Monitor Security at an Industrial Facility

The article deals with the tasks of searching and classifying objects in images for the purpose of analyzing industrial risks. The objects of recognition are the presence of signal vests and protective helmets for an employee of production facilities. The YOLOv3 (You Only Look Once) convolutional neural network is used as a research tool. Unlike other convolutional neural networks, YOLO only needs to “look” at the image once to detect objects. This significantly reduces the time required to detect an object in the frame, while maintaining recognition accuracy. The practical implementation was made in the Python programming language for the Windows platform. The article contains a description of the stages of training a neural network and assessing the quality of the resulting model. Training and test sets (dataset) were formed in the Pascal VOC format, containing images and their annotations in the form of an XL file. During training, “transfer irradiation” was used. It involves using as a starting point the weight from a pre-trained model on the COCO dataset. As a result of training, 30 models were obtained. To assess the quality of work of each neural network model obtained during training, a test data set was used. A summary table of the results of testing the neural network model is also provided, which contains the average accuracy for each class (vest class - vest and helmet class - helmet), as well as the mAP metric. The result of the work of the trained YOLOv3 convolutional neural network model is an image marked with rectangles with labels to which class the recognized object belongs and with what accuracy it belongs to this class. Based on the results obtained, it is possible to build higher-level models for use in research purposes, as well as to create systems for analyzing and monitoring safety at production facilities.