Original scientific paper

Surface defect detection of steel based on improved YOLOv7 model

W. Z. Teng ; School of Computer Science and Software Engineering, University of Science and Technology Liaoning, China
Y. J. Zhang ; School of Computer Science and Software Engineering, University of Science and Technology Liaoning, China *
H. G. Zhang ; School of Computer Science and Software Engineering, University of Science and Technology Liaoning, China
D. X. Gao ; School of Computer Science and Software Engineering, University of Science and Technology Liaoning, China

* Corresponding author.

Abstract

In response to the inevitable surface defects in the manufacturing process of hot-rolled steel, this paper proposes an improved steel surface defect detection model based on YOLOv7. In the Extended Efficient Large Aggregation Network (E-ELAN), the model replaces conventional convolution with Omni-Dimensional Dynamic Convolution (ODConv) to enhance the network’s sensitivity to feature extraction using a combination of various attention mechanisms. Additionally, the detection head in the head section is replaced with an Efficient Decoupled Detection Head, enhancing the model’s capability to classify and locate small defects. The proposed model is tested on the public dataset NEU-DET, achieving a high mAP of 76,5 %. This effectively enhances the model’s ability to detect surface defects in steel while maintaining a fast detection speed.

Keywords

hot-rolled; steel; surface defect; YOLOv7 model; efficient decoupled detection head

Hrčak ID:

315685

URI

https://hrcak.srce.hr/315685

Publication date:

1.7.2024.

Visits: 906 *