ELA-YOLO: An efficient method with linear attention for steel surface defect detection during manufacturing

Abstract

Research on deep learning methods for steel surface defect detection significantly enhances product quality and manufacturing efficiency. However, practical industrial scenarios pose challenges, including variations in color, lighting, reflective conditions, and other environmental factors that affect defect visibility. Additionally, defects vary in size and shape, with some being so small or concealed that accurate detection is difficult. Complex textures of detected images further increase computational cost, often compromising efficiency for high precision. In this paper, we propose a novel method called ELA-YOLO for defect detection, using YOLOv8 as the underlying framework. First, we introduce linear attention to the network to improve the model’s representation capability while managing computational complexity. Second, we propose a selective feature pyramid network to enhance feature fusion across different levels. Third, we design a lightweight detection head to output detection results efficiently. Experimental results demonstrate that ELA-YOLO achieves the highest accuracy: 81.7 mAP on the NEU-DET dataset, 99.3 mAP on the DAGM2007 dataset and 74.3 mAP on the GC10-DET dataset. Additionally, it achieves the lowest parameters (5.4 M), computational complexity (16.5 GFLOPs), and relatively low latency (101.3 FPS). Our method strikes an optimal balance between efficiency and accuracy, demonstrating comprehensive performance in industrial steel surface defect detection.