Detection of Defects in Industrial Pipelines Based on Computer Vision: A Comparative Study

De Lama Jimenez A.; Carbaial Balarezo R.; Inga Espinoza C. · Proceedings of the 2025 IEEE 32nd International Conference on Electronics, Electrical Engineering and Computing, INTERCON 2025 (2025) · DOI: 10.1109/intercon67304.2025.11244661

Resumen

This work addresses the detection of internal faults in industrial pipelines, specifically cracks, dents, and scratches. Two deep learning models, YOLOv8 and U-Net, were trained and evaluated for this purpose. The dataset consisted of real-world images capturing various types of pipeline damage. The performance comparison focused on detection accuracy, classification capability, inference speed, and computational efficiency within the context of industrial inspection. Experimental results indicate that YOLOv8 achieved an inference speed of 13-14 frames per second, with an overall precision of 82.5 % and an Intersection over Union (IoU) of 75.9% for the scratch class. In contrast, U - N et reached a precision of 91.3 %, a Dice coefficient of up to 88.9%, and an inference speed of 1-2 frames per second.

Instituciones: Universidad Peruana de Ciencias Aplicadas