| Peng Siyu (彭思雨)* **,Li Dingxin*,Wang Huan*,Li Fang*,TIAN Dongping*,Shi Zhongzhi***.[J].高技术通讯(英文),2026,32(1):84~96 |
|
| Research on steel surface detection based on YOLOv5s-SNTC model |
| |
| DOI：10. 3772 / j. issn. 1006-6748. 2026. 01. 009 |
| 中文关键词: |
| 英文关键词: YOLOv5s, steel surface, object detection, convolutional block attention module |
| 基金项目: |
| Author Name | Affiliation | | Peng Siyu (彭思雨)* ** | (* Institute of Information Engineering, Changji University, Changji 831100, P. R. China)
(** Key Laboratory of Computer Application Technology, Changji University, Changji 831100, P. R. China)
(*** Key Laboratory of Intelligent Information Processing, Chinese Academy of Sciences, Beijing 100190, P. R. China) | | Li Dingxin* | | | Wang Huan* | | | Li Fang* | | | TIAN Dongping* | | | Shi Zhongzhi*** | |
|
| Hits: 23 |
| Download times: 29 |
| 中文摘要: |
| |
| 英文摘要: |
| Steel surface detection is a critical component in ensuring product quality. Traditional manual and photoelectric inspection methods suffer from low efficiency, high costs and high false detection rates. To address these issues, this paper presents a lightweight, high-precision steel surface detection model YOLOv5s (you only look once version 5 small)-SNTC ( short for Slim-Neck with three CBAM modules). To start with, the CBAM (short for convolutional block attention module) module is integrated into the backbone to enhance feature representation capability. Subsequently, the neck is replaced with a Slim-Neck structure to improve feature fusion performance while significantly reducing the model’s parameter count and computational complexity. Furthermore, the model’ s network structure is meticulously designed with computational resources allocated optimally across layers to ensure the model sustaining lightweight properties while fully exerting its detection performance.Finally, extensive experiments conducted on the NEU-DET (NEU surface defect database) dataset validate that the YOLOv5s-SNTC model achieves a mean average precision (mAP) of 76. 6% and a detection speed of 155 frames per second. Compared to the baseline model YOLOv5s, the mAP of the YOLOv5s-SNTC model is significantly improved by 7. 6% . Meanwhile, its model size is compressed to 11. 6 MB, and its computational load is reduced to 13. 9 giga floating-point operations per second (GFLOPS) . This demonstrates that the model proposed here can effectively improve the detection speed and accuracy in steel surface detection without significantly increasing complexity and resource consumption, achieving an optimal balance between lightweight design and high precision. |
|
View Full Text
View/Add Comment Download reader |
| Close |
|
|
|
