Year: 2025
Type: article
Abstract: The research is devoted to the analysis and development of methods for generating artificial data in order to solve computer vision problems that arise during the operation of conveyor-type technological lines. The paper substantiates the relevance of this problem, as traditional approaches to data collection, such as video recording and manual markup, are not only time-consuming but also ineffective in terms of cost and efficiency. These methods are often not applicable in situations where a large amount of data needs to be processed or in cases where efficiency is required. As an alternative approach, the paper proposes a generalized model for synthetic data generation using modern game engines and 3D modeling techniques. This model allows for a high degree of variability and realism in the generated data, making it possible to simulate various scenarios for conveyor lines and specific computer vision tasks. The effectiveness of the proposed approach was experimentally confirmed using the example of the problem of determining laminate coatings. Synthetic data generated using the proposed model were used to train the YOLOv8 neural network as part of the experiment. The test results showed a high accuracy of the trained model, with an mAP50 of 0.95, indicating the significant potential of synthetic data for improving the quality of machine learning models. These results confirm the possibility of using synthetic data when access to real datasets is limited. This opens up opportunities for optimizing neural network learning processes and improving the efficiency of computer vision solutions, particularly in industrial applications.
Authors: A. V. Zaripov, R. S. Kulshin, A. A. Sidorov
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $500
Award: G2206865297
OpenAlex ID: W4409568526