Performance Comparison and Analysis of YOLOv8/v9/v10 in Steel Plate Surface Defect Detection

Abstract

During steel plates’ continuous casting and rolling, surface defects often arise from rolling equipment/processes, risking production accidents. These defects are small, easily occluded, and require real-time detection for efficiency—demanding high accuracy and speed from detection algorithms. This paper synthesizes relevant research papers on the application of You Only Look Once version 8, version 9, version 10 models (YOLOv8, v9, v10 models), and their improved versions in steel surface quality inspection. It analyzes and summarizes the technological evolution path in the field of industrial quality inspection, with a key focus on the application and performance of the YOLO series models. Studies indicate that single-stage detection models have gained significant attention due to their advantage in balancing speed and accuracy. Among them, YOLO models and their improved variants have become the industry mainstream through a series of enhancements, such as architecture decoupling and feature compression. Specifically, YOLOv10 and its improved models, benefiting from innovations in non-NMS inference, dynamic gate-controlled feature fusion, shape awareness, and loss functions, have achieved a detection accuracy of 85.5% mAP@0.5 on datasets like NEU-DET. This study will further summarize the limitations of current technologies in small-sample dense defect recognition and propose development directions centered on multi-modal fusion and adaptive computing.