Simulation of a neural network algorithm for the analysis of video flow defects

Abstract

The paper provides an overview of existing scientific works on video analysis for defect detection and digital broadcasting technologies. The creation of a data set, data distribution by samples are considered. The training of neural network models and obtaining training re-sults are described. Implementation of the method of using a trained neural network model. From the results obtained, we can draw an unfortunate conclusion that the trained model YOLOv8x.pt has the highest accuracy in detecting defects. After obtaining the best results of model training, it was found that the highest resolu-tion affects the quality of error detection only up to a certain value. On this training sample, a sufficient resolution for image detection is 540x540 pixels. The computing resources of the Google Colaboratory service are limited for free use, so the number of training epochs was reduced for the YOLOv8x.pt and yolov5x6u.pt models. Even with fewer training epochs, the YOLOv8x.pt model was able to show the best re-sult. For further use, the trained neural network model YOLOv8x.pt will be used.