Steel Defect Detection Based on YOLO-SAFD

Abstract

In industrial production, defect detection of steel materials is critical for maintaining quality. However, traditional inspection methods are labor-intensive and error-prone, while existing deep learning-based detection approaches generally suffer from poor performance in industrial defect detection and insensitivity to small defects. This paper presents YOLO-SAFD, an advanced framework based on YOLOv5, designed to address these challenges. The proposed model incorporates two key innovations: 1) the Squeezed and Excited Asymptotic Feature Pyramid Network (SAFPN), which enhances multi-scale feature fusion and improves the detection of small defects, increasing the mean Average Precision (mAP) from 0.78 (YOLOv5 baseline) to 0.84; 2) the Diverse Branch Block (DBB), which replaces conventional convolutions to enrich feature diversity while reducing computational complexity, cutting the model parameters from 13.8M to 4.82M. Experimental results on the NEU-DET dataset demonstrate that YOLO-SAFD achieves a detection precision of 0.83, a recall of 0.75, and an mAP50:95 of 0.43, outperforming the baseline YOLOv5 and highlighting its superior detection accuracy and efficiency for real-time industrial applications.