YOLO-VMTC: Enhancing Steel Surface Defect Detection with a Lightweight and Context-Aware Deep Learning Approach

Abstract

In the steel manufacturing process, the surface quality of steel not only reflects the integrity of the steel surface but also significantly influences the quality and safety of downstream production equipment. To address the shortcomings of traditional steel surface defect detection models, such as insufficient feature extraction capability, low detection accuracy, and the large computational load and model size, we propose an optimized YOLOv8-based algorithm for steel surface defect detection, termed YOLO-VMTC, which incorporates lightweight and context-aware features. The stated detection model augments the conventional YOLOv8 backbone architecture through the fusion of VanillaNet and the Multi-Scale Contextual Feature module. This integration upholds the backbone's lightweight design while incorporating multi-scale contextual features, consequentially bolstering the model's feature extraction prowess. Additionally, the neck network of YOLOv8 undergoes refinement via the introduction of the Triplet Attention Mechanism and C2fSKConv module. This adjustment minimizes computational demands and mitigates interference from irrelevant information, empowering the model to precisely identify and emphasize defect locations on steel surfaces. Furthermore, the substitution of the conventional IoU loss with WiseIoU loss within the detection framework accelerates training, conserving both computational power and time. The inclusion of Meta-ACON, a component capable of adaptively modulating its activation patterns, fortifies the model's generalization capacity, particularly when confronted with intricate and dynamic tasks like steel surface defect analysis. Experimental evaluations on the NEU-DET steel surface defect dataset reveal that this model attains an impressive mAP of 79.0% while maintaining a swift detection speed of 96.3 FPS. When compared to the baseline YOLOv8 defect detection model, it exhibits enhancements not only in detection accuracy but also in speed, offering a robust foundation for quality assurance within steel manufacturing processes.