Detection of Surface Defects in Metallic Materials Using Convolutional Neural Networks with YOLO Architecture

Read full text Summarise with VIRAT
Authors: Roman Bumbálek, Tomáš Zoubek, Ján Majerník, Jean de Dieu Marcel Ufitikirezi, Sandra Nicole Umurungi, Karel Šramhauser, František Špalek
Cited by: 0
References: 1

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

The early detection and classification of surface defects in metallic materials is essential for ensuring product quality and reliability in industrial production. In this study, we evaluated and compared five recent versions of the YOLO convolutional neural network architecture (YOLOv8x, YOLOv9e, YOLOv10x, YOLOv11x, and YOLOv12x) applied to five publicly available datasets specialized in metallic defect detection (AlcastXray, Castings, GC10-DET, NEU-SEG, and Severstal SDD). All models were trained under uniform conditions and assessed using Precision, Recall, mAP50, mAP50–95, and deployment-oriented efficiency metrics, including inference speed, throughput, GPU memory usage, and power draw. The results show that YOLOv9e consistently achieved the highest detection accuracy (mAP50 up to 0.982), while YOLOv8x provided the fastest inference (>47 FPS), making it suitable for real-time applications. YOLOv10x, the most lightweight model, delivered a favorable trade-off between accuracy and computational efficiency, suggesting strong potential for edge deployment. These findings provide practical insights for selecting YOLO-based architectures according to specific industrial requirements in metallic surface defect inspection.

Fields of Study & Subjects

Industrial Vision Systems and Defect Detection Advanced Neural Network Applications Non-Destructive Testing Techniques

Publication details

Journal / source
Archives of Foundry Engineering
Pages
165–176
Publication date
30 March 2026
Language
PL