An accurate, generalized, and efficient detection framework for steel surface defect inspection

Abstract

Steel surface defect detection is crucial for ensuring product quality and process reliability in industrial manufacturing; however, existing methods often struggle to simultaneously achieve high accuracy, real-time efficiency, and strong cross-dataset generalization. This paper proposes AGE-YOLO (Accurate, Generalized, and Efficient YOLO (You Only Look Once)), an enhanced detection framework built upon the YOLO version 11 (YOLOv11) architecture. It is designed to address these challenges by incorporating three key components: (1) a Feature-Context Multi-Aggregation Network (FC-MANet) that integrates Convolutional Gated Linear Units (CGLU) and Partial Convolution (PConv) to capture irregular defect features while mitigating background artifacts through dynamic gating; (2) an Adaptive Cross-scale Fusion (ACF) module that dynamically aligns and fuses multi-level features via joint spatial–channel guidance, effectively suppressing irrelevant background responses; and (3) a Shared Insight Focus Head (SIFHead) employing lightweight shared convolutions with Group Normalization (GN) to enhance localization consistency, particularly for small defects.

Experimental results on the GC10-DET dataset demonstrate that AGE-YOLO achieves a mean Average Precision (mAP) of 72.8% at an Intersection over Union (IoU) threshold of 0.5 ([email protected]), outperforming YOLOv11n (nano) baseline by 5.2% and recent YOLOv12n and YOLOv13n variants by 5.8% and 5.6%, respectively. The method also surpasses the Real-Time DEtection TRansformer (RT-DETR) by 3.6% mAP while using only 10.4% of its parameters. With 3.41 million parameters and 7.7 Giga Floating-Point Operations (GFLOPs), AGE-YOLO achieves a real-time inference speed of 189 frames per second (FPS) at 640 × 640 resolution on an NVIDIA RTX 4090D Graphics Processing Unit (GPU). Furthermore, hardware validation on the NVIDIA Jetson Orin Nano platform shows that, after optimization with the NVIDIA TensorRT inference engine using 16-bit floating-point (FP16) precision, the model attains a throughput of 74.6 FPS, confirming its suitability for real-time edge deployment. Additional experiments on the Northeastern University steel defect dataset (NEU-DET) and the printed circuit board defect dataset released by the Peking University Open Laboratory on Human–Robot Interaction (HRIPCB) demonstrate [email protected] scores of 85.4% and 95.7%, respectively, yielding an average improvement of 4.5% over representative lightweight detectors. The implementation of this work is publicly available at https://github.com/gnmtdt/AGE-YOLO.