TITLE:
A YOLOv8-Based Network for Surface Defect Detection in Strip Steel
AUTHORS:
Jiacheng Jiang, Yuebin Su, Jing Song
KEYWORDS:
Lightweight Network, Starnet, Self-Concatenated Feature Enhancement, Surface Defect Detection, YOLOv8
JOURNAL NAME:
Journal of Computer and Communications,
Vol.14 No.2,
February
14,
2026
ABSTRACT: Effectively detecting subtle surface defects in strip steel is vital for industrial quality assurance; however, most existing approaches fail to strike an optimal balance between accuracy and efficiency, especially under real-time processing constraints. To address these limitations, we propose SGS-YOLO, a lightweight detection network built upon an enhanced YOLOv8 architecture. In this work, we offer three core contributions through our proposed method. First, we incorporate the star operation into YOLOv8’s backbone and C2f modules to reduce the model’s parameters and computational cost. Second, we introduce a Group-Sharing Detection Head (GS-Detect), leveraging Group Normalization and shared convolutions to jointly enhance accuracy and inference efficiency. Third, we propose a novel Self-Concatenated Feature Enhancement (SCFE) mechanism to enrich feature diversity and gradient propagation. Notably, SCFE itself is parameter-free, while the reported overall parameter/FLOPs increase arises from subsequent layers that receive widened inputs after concatenation. Compared with YOLOv8n, SGS-YOLO improves mAP@50 by 2.9%, respectively, on the NEU-DET dataset. With 1.49 M parameters and an inference speed of 130.22 FPS, SGS-YOLO achieves a favorable accuracy-efficiency trade-off. Compared to other mainstream models, SGS-YOLO demonstrates strong detection accuracy while maintaining a lightweight architecture, which satisfies the quasi-real-time requirement in multiscale defect detection task.