Abstract
<jats:p>In the context of increasing wear and tear of utility pipelines, improving the quality of diagnostics of their technical condition, as well as improving the quality of control over construction and repair work, is of particular importance. Timely detection of defects and damage helps prevent emergency situations and negative socio-economic and environmental consequences. Traditional methods of visual and instrumental control have a number of disadvantages, in particular, high labor intensity, long information processing time, and insufficient accuracy. This problem can be solved by using computer vision to identify and classify damage, which will improve the quality of defect detection, reduce the likelihood of human error, and speed up the diagnostic process. However, different levels of readiness of the developed machine learning algorithms require additional research to confirm the effectiveness of their use in professional fields, for example, when inspecting structures, which justifies the relevance of this study. The object of the study was the YOLO family of computer vision models capable of identifying various classes of defects. The aim of the study was to train the YOLOv5, YOLOv8, and YOLO11 detection algorithms and to perform a comparative analysis of their speed and accuracy of data processing using weld defects as an example. The results of the experimental studies show that the use of the latest version of the model does not lead to significant improvements in the quality of defect detection compared to previous versions. The results presented in the work allow us to assess the feasibility of using the new YOLO11 model to detect defects in radiographic images. Based on the experiments, researchers using computer vision methods to control the quality of welds can make an informed decision about whether to use this model or use previous versions of the algorithms.</jats:p>