FEP-YOLO: a lightweight steel surface defect detection method for resource-constrained devices

Abstract The detection of surface defects in steel is a critical task essential for ensuring the quality of industrial products. Deep learning-based defect detection research has demonstrated significant advancements and promising results. However, deep learning-driven detection methods generally rely on high-performance computing hardware for implementation, which restricts their applicability in resource-constrained environments. To resolve these challenges, this study designs a lightweight detection framework, termed FasterNet, efficient multi-scale attention (EMA) mechanism, partial convolution detect (PC-Detect)-you only look once, specifically designed for resource-constrained devices. Firstly, a streamlined fundamental network structure, FasterNet, is adopted to substantially reduce the model size while preserving detection accuracy. Secondly, an EMA mechanism is integrated in the lower layers of the feature extraction network to enhance the model’s ability to accurately locate defects. Lastly, a lightweight detection head, PC-Detect, is developed to further minimize redundant computations and reduce model complexity. Experimental results demonstrate that the proposed model achieves a mean average precision of 80.9% on the NET-DET dataset, with a detection speed of 31.8 frames per second on a CPU. To assess the enhanced model performance on resource-constrained devices, it was deployed on a Raspberry Pi 5, achieving an inference speed of 55 ms. This study validates the feasibility of performing high-quality defect detection on resource-constrained devices, providing low-cost quality monitoring system for industrial manufacturing and broadening the scope of edge computing applications within the industrial sector.