An Improved YOLOv8 Model for Strip Steel Surface Defect Detection DOI Creative Commons
Jinwen Wang,

Ting Chen,

Xinke Xu

et al.

Applied Sciences, Journal Year: 2024, Volume and Issue: 15(1), P. 52 - 52

Published: Dec. 25, 2024

In the process of steel strip production, accuracy defect detection remains a challenge due to diversity types, complex backgrounds, and noise interference. To improve effectiveness surface in strips, we propose an enhanced model known as YOLOv8-BSPB. First, novel pooling layer module, SCRD, which replaces max with average pooling. This module introduces receptive field block (RFB) deformable convolutional network version 4 (DCNv4) obtain learnable offsets, allowing kernels flexibly move deform on input feature map, thus, more effectively extracting multi-scale features. Second, integrate polarized self-attention (PSA) mechanism model’s representation enhance its ability focus relevant information. Additionally, incorporate BAM attention after C2f strengthen selection capabilities. A bidirectional pyramid is introduced at neck transmission efficiency. Finally, WIoU loss function employed accelerate convergence speed regression accuracy. Experimental results NEU-DET dataset demonstrate that improved achieves classification 81.3%, increase 4.9% over baseline, mean precision 86.9%. The has parameter count 5.5 M operates 103.1 FPS. validate effectiveness, conducted tests Kaggle our custom dataset, where by 2.3% 5.5%, respectively. experimental indicate meets requirements for real-time, lightweight, portable deployment.

Language: Английский

An Improved YOLOv8 Model for Strip Steel Surface Defect Detection DOI Creative Commons
Jinwen Wang,

Ting Chen,

Xinke Xu

et al.

Applied Sciences, Journal Year: 2024, Volume and Issue: 15(1), P. 52 - 52

Published: Dec. 25, 2024

In the process of steel strip production, accuracy defect detection remains a challenge due to diversity types, complex backgrounds, and noise interference. To improve effectiveness surface in strips, we propose an enhanced model known as YOLOv8-BSPB. First, novel pooling layer module, SCRD, which replaces max with average pooling. This module introduces receptive field block (RFB) deformable convolutional network version 4 (DCNv4) obtain learnable offsets, allowing kernels flexibly move deform on input feature map, thus, more effectively extracting multi-scale features. Second, integrate polarized self-attention (PSA) mechanism model’s representation enhance its ability focus relevant information. Additionally, incorporate BAM attention after C2f strengthen selection capabilities. A bidirectional pyramid is introduced at neck transmission efficiency. Finally, WIoU loss function employed accelerate convergence speed regression accuracy. Experimental results NEU-DET dataset demonstrate that improved achieves classification 81.3%, increase 4.9% over baseline, mean precision 86.9%. The has parameter count 5.5 M operates 103.1 FPS. validate effectiveness, conducted tests Kaggle our custom dataset, where by 2.3% 5.5%, respectively. experimental indicate meets requirements for real-time, lightweight, portable deployment.

Language: Английский

Citations

1