[1]蔡盼盼,刘 娟,鲁忠臣.基于 EfficientNetB3 与嵌入式系统的热轧钢带表面缺陷分拣系统[J].机械与电子,2026,44(04):62-67.
　CAI Panpan,LIU Juan,LU Zhongchen.A Surface Defect Sorting System for Hot-rolled Steel Strip Based on EfficientNetB3 and Embedded System[J].Machinery & Electronics,2026,44(04):62-67.
点击复制

基于 EfficientNetB3 与嵌入式系统的热轧钢带表面缺陷分拣系统()
分享到：

《机械与电子》[ISSN:1001-2257/CN:52-1052/TH]

卷:
44
期数:
2026年04期
页码:
62-67
栏目:
智能检测
出版日期:
2026-04-27

文章信息/Info

Title:
A Surface Defect Sorting System for Hot-rolled Steel Strip Based on EfficientNetB3 and Embedded System
文章编号:
1001-2257 ( 2026 ) 04-0062-06
作者:
蔡盼盼; 刘 娟; 鲁忠臣
华南理工大学工程训练中心,广东 广州 510641
Author(s):
CAI Panpan ;  LIU Juan ;  LU Zhongchen
( Engineering Training Center , South China University of Technology , Guangzhou 510641 , China )
关键词:
表面缺陷检测; EfficientNetB3 ; 嵌入式系统; 实时分拣
Keywords:
surface defect detection ;  EfficientNetB3 ;  embedded system ;  real-time sorting
分类号:
TP391.4
文献标志码:
A
摘要:
针对工业场景下钢带表面缺陷检测人工检测效率低、传统机器视觉方法泛化能力差的问题,提出一种融合深度学习与嵌入式控制的自动化分拣系统。在算法层面,构建了以 EfficientNetB3 为骨干的模型,通过 2 个阶段微调策略优化,在 NEU-DET 数据集上实现了 0.99 的平均分类精确率,单张图像推理时间约 58 ms ,满足了工业实时性要求。在系统层面,设计并搭建了上位机 下位机构成的协同硬件平台,集成了图像采集、缺陷识别和机械分拣等多个功能模块。集成测试结果表明,该系统对斑块、氧化皮、裂纹、划痕、点蚀和夹杂物 6 类典型缺陷的整体分拣成功率达 97.8% ,有效验证了其在工业现场实现高精度、高效率自动化质检的应用潜力与实用价值。
Abstract:
To address the issues of low efficiency in manual detection and poor generalization capability of traditional machine vision methods for surface defects on steel strips in industrial scenarios , an automated sorting system that integrates deep learning with embedded control is proposed.At the algorithmic level , a transfer learning model based on the EfficientNetB3 backbone is constructed and optimized via a two stage fine tuning strategy.This model achieves an average classification precision of 0.99 on the NEU DET dataset , with a single image inference time of approximately 58 ms , satisfying industrial real-time requirements.At the system level , a collaborative hardware platform comprising an upper computer ( decision making ) and a lower computer ( execution ) was designed and implemented , integrating multiple functional modules such as image acquisition , defect recognition , and mechanical sorting.Integrated test results demonstrate that the system achieves an overall sorting accuracy of 97.8% for six typical defect types : patches , rolled in scale , cracks , scratches , pitted surfaces , andinclusions.This effectively validates its application potential and practical value for achieving high-precision and high-efficiency automated quality inspection in industrial settings.

参考文献/References:

[ 1 ] 马天龙,李学伟,丁峰,等 . 基于改进 YOLOv5 算法的齿轮表面缺陷检测方法[ J ] . 组合机床与自动化加工技术,2025 ( 4 ): 152-156.

[ 2 ] 王震洲,杨榕,宿景芳 . 改进YOLOv8的烟盒缺陷检测算法[ J ] . 电子测量技术, 2024 , 47 ( 13 ): 110-119.
[ 3 ] 冒浩杰,巩永旺 . 改进YOLOv5s的小目标钢材表面缺陷检测算法[ J ] . 电子科技, 2025 , 38 ( 10 ): 10-18.
[ 4 ] SIMONYAN K , ZISSERMAN A.Very deep convolutional networks for large-scale image recognition [ EB / OL ] . [ 2025-11-30 ] .2014.https : ∥doi.org / 10.48550 / arXiv.1409.1556.
[ 5 ] HE K M , ZHANG X Y , REN S Q , et al.Deep residual learning for image recognition [ C ] ∥Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition , 2016 : 770-778.
[ 6 ] KING P R , POPESCU A T , HARKIN T G.Enabling on-device CNN customization via in-situ updates and unsupervised data [ J ] .ACM Transactions on embedded computing systems , 2024 , 23 ( 3 ): 1-25.
[ 7 ] BI S , WANG S , WANG C , et al.A multi-objective evolutionary architecture pruning of DCNN with weight inheritance [ J ] .Information sciences , 2024 , 658 : 120026.
[ 8 ] LI J , WANG M , ZHANG C.A lightweight neural architecture search for efficient model deployment on edge devices [ J ] .IEEE Transactions on industrial informatics , 2023 , 19 ( 10 ): 10145-10155.
[ 9 ] SMITH A , JOHNSON B.Challenges and optimizations for deploying deep neural networks like ResNet on resource-constrained edge AI processors [ C ] ∥Proceedings of the IEEE International Conference on Edge Computing ( EDGE ), 2024 : 123-130.
[ 10 ] WANG J , LI X , ZHANG Y , et al.A lightweight network for road damage detection leveraging coordinate attention and feature fusion mechanisms [ J ] .Informa- tion fusion , 2024 , 108 : 102405.
[ 11 ] CHEN Y , WU Z , LI H , et al.Fabric defect detection based on asymmetric dual-stream network and deformable convolution [ J ] .Engineering applications of artificial intelligence , 2023 , 126 : 107149.
[ 12 ] 帖军,隆娟娟,郑禄,等 . 基于 SK-EfficientNet 的番茄叶片病害识别模型[ J ] . 广西师范大学学报(自然科学版),2022 , 40 ( 4 ): 104-114.
[ 13 ] 马子儒,刘云鹏,裴少通,等 . 基于轻量级卷积神经网络的复合绝缘子憎水性诊断方法研究[ J ] . 高压电器,2024 , 60 ( 3 ): 93-100.
[ 14 ] PARMAR V , BHATIA N , NEGI S , et al.Exploration of optimized semantic segmentation architectures for edge-deployment on drones [ C ] ∥2023 IEEE India Geoscience and Remote Sensing Symposium , 2024 :1-4.

相似文献/References:

[1]梁 程,薛建彬.基于云-边缘协同计算的表面缺陷检测系统研究[J].机械与电子,2022,(02):65.
　LIANG Cheng,XUE Jianbin.Research on a Surface Defect Detection System Based on the Cloud-edge Computing[J].Machinery & Electronics,2022,(04):65.
[2]肖 红,陈希球.基于图像角点匹配的机械加工零件表面缺陷检测[J].机械与电子,2022,(08):62.
　XIAO Hong,CHEN Xiqiu.Surface Defect Detection of Machined Parts Based on Image Corner Matching[J].Machinery & Electronics,2022,(04):62.

备注/Memo

备注/Memo:
收稿日期: 2025-12-30
基金项目: 2025 年度广东省本科高校教学质量与教学改革工程建设项目(粤教高函〔 2026 〕 4 号)
作者简介:蔡盼盼 ( 1988- ),女,江苏徐州人,硕士,实验师,研究方向为深度学习与机器视觉。
更新日期/Last Update: 2026-08-20