What type of study is this?

This is a Quantitative Study study.

September 16, 2025

ASD-YOLO: A Lightweight Multi-Module Collaboratively Optimized Model for Steel Surface Defect Detection

WFWeiyue FangWenzhou Medical University YYYong YangGuangdong Polytechnic Normal University WZWang ZhangGuangdong Polytechnic Normal University

Puntos clave

ASD-YOLO achieves a mean Average Precision (mAP) of 77.3% on the NEU-DET dataset, and enhances performance over the YOLO11 baseline.
The model reduces parameters by 3.5% while improving computational efficiency, crucial for real-time defect detection.
Multi-module optimization addresses challenges in small-object modeling, integrating advanced functionalities for better feature transmission.
The model demonstrates strong generalization on the GC10-DET dataset, suggesting applicability across various industrial contexts.

Resumen

Abstract Surface defect detection of steel is a critical step in ensuring the quality and safety of industrial products. In particular, detecting small-scale defects such as scratches and spots—characterized by blurred edges and low contrast—poses stringent requirements for both the accuracy and real-time performance of detection algorithms. To address the limitations of traditional methods—namely low efficiency and high cost—as well as the shortcomings of existing deep learning models in detecting small objects with high accuracy and low computational cost, this paper proposes an improved lightweight object detection model for steel surface defect detection, termed ASD-YOLO. Built upon the YOLO11 architecture, the model conducts coordinated multi-module optimization around three key aspects: small-object modeling, feature fusion, and structural lightweighting. The backbone integrates the ACmix module, which combines local convolution and global attention to enhance detail perception capabilities. A Slim-Neck module is employed in the neck to reduce parameter size while improving feature transmission efficiency. The Shape-IoU loss function is introduced to improve the model's ability to represent irregular defect boundaries. A DLGCA attention mechanism is incorporated prior to the detection head, significantly enhancing the model’s discrimination capability in complex textured backgrounds. Experimental results show that ASD-YOLO achieves a mAP@0.5 of 77.3% on the NEU-DET dataset, marking a 5.4% improvement over the baseline YOLO11 model. The model also achieves a 3.5% reduction in parameters and a 7.9% decrease in computational cost, while demonstrating strong generalization performance on the GC10-DET dataset. These results indicate that the proposed method achieves synergistic improvements in detection accuracy, model compactness, and industrial applicability, offering a reliable and efficient solution for steel defect detection.

Preguntar a la IA

Me gusta

Guardar

Ver artículo completo