SC-AttentiveNet: Lightweight Multiscale Feature Fusion Network for Surface Defect Detection on Copper Strips DOI Open Access
Zeteng Li, Guo Zhang,

Qi Yang

et al.

Electronics, Journal Year: 2025, Volume and Issue: 14(7), P. 1422 - 1422

Published: April 1, 2025

Small defects on the surface of copper strips have a significant impact key properties such as electrical conductivity and corrosion resistance, existing inspection techniques struggle to meet demand in terms accuracy generalisability. Although there been some studies metal defect detection, is relative lack research highly reflective strips. In this paper, lightweight efficient strip detection algorithm, SC-AttentiveNet, proposed, aiming solve problems large model size, slow speed, insufficient poor generalisability models. The algorithm based ConvNeXt V2, combines SCDown module group normalisation design SCGNNet feature extraction network, which significantly reduces computational overhead while maintaining excellent capability. addition, introduces SPPF-PSA enhance multi-scale capability, constructs new neck fusion network via HD-CF Fusion Block module, further enhances diversity fine granularity. experimental results show that SC-AttentiveNet has mAP 90.11% 64.14% KUST-DET VOC datasets, respectively, with parameter count only 6.365 MB complexity 14.442 GFLOPs. Tests NEU-DET dataset an generalisation performance, 76.41% speed 78 FPS, demonstrating wide range practical application potential.

Language: Английский

SC-AttentiveNet: Lightweight Multiscale Feature Fusion Network for Surface Defect Detection on Copper Strips DOI Open Access
Zeteng Li, Guo Zhang,

Qi Yang

et al.

Electronics, Journal Year: 2025, Volume and Issue: 14(7), P. 1422 - 1422

Published: April 1, 2025

Small defects on the surface of copper strips have a significant impact key properties such as electrical conductivity and corrosion resistance, existing inspection techniques struggle to meet demand in terms accuracy generalisability. Although there been some studies metal defect detection, is relative lack research highly reflective strips. In this paper, lightweight efficient strip detection algorithm, SC-AttentiveNet, proposed, aiming solve problems large model size, slow speed, insufficient poor generalisability models. The algorithm based ConvNeXt V2, combines SCDown module group normalisation design SCGNNet feature extraction network, which significantly reduces computational overhead while maintaining excellent capability. addition, introduces SPPF-PSA enhance multi-scale capability, constructs new neck fusion network via HD-CF Fusion Block module, further enhances diversity fine granularity. experimental results show that SC-AttentiveNet has mAP 90.11% 64.14% KUST-DET VOC datasets, respectively, with parameter count only 6.365 MB complexity 14.442 GFLOPs. Tests NEU-DET dataset an generalisation performance, 76.41% speed 78 FPS, demonstrating wide range practical application potential.

Language: Английский

Citations

0