文献类型：期刊文献

中文题名：基于ECA-ResNet的热轧钢表面缺陷在线识别

英文题名：Online Identification of Hot-rolled Steel Surface Defects Based on ECA-ResNet

作者：杨子辉[1];刘艳霞[1]

第一作者：杨子辉

机构：[1]北京联合大学城市轨道交通与物流学院,北京100101

第一机构：北京联合大学城市轨道交通与物流学院

年份：2024

卷号：38

期号：2

起止页码：59-65

中文期刊名：北京联合大学学报

外文期刊名：Journal of Beijing Union University

基金：北京联合大学科研项目(ZK20202302),北京市自然基金项目(L221015)。

语种：中文

中文关键词：注意力机制;ResNet;热轧钢;图像分类

外文关键词：attention mechanism;ResNet;hot-rolled steel;image classification

摘要：基于视觉注意力机制对ResNet模型进行改进,设计了ECA-ResNet网络模型,对热轧钢表面缺陷进行在线识别。该模型在提高检测精度的同时,其推理速度也符合热轧钢生产线对实时性的要求。实验结果表明,ECA-ResNet模型的在线识别精度为98.1%,比经典ResNet网络提高了3.9个百分点。消融实验和对比实验说明,融合ECA注意力模块的ResNet网络的综合性能优于融合CBAM或SE注意力模块的ResNet网络,也优于经典ResNet网络、VGG-16网络和GoogLeNet网络,符合热轧钢表面缺陷的在线识别要求,具有良好的应用前景。
This article improves the ResNet model based on visual attention mechanisms and designs the ECA-ResNet network model for online identification of hot-rolled steel surface defects.While enhancing the detection accuracy,the model’s inference speed also meets the real-time requirements of the hot-rolled steel production line.Experimental results show that the online recognition accuracy of the ECA-ResNet model is 98.1%,which is 3.9 percentage points higher than the classic ResNet network.Ablation experiments and comparative experiments demonstrate that the ResNet network integrated with the ECA attention module outperforms those integrated with the CBAM or SE attention modules and is superior to the classic ResNet network,VGG-16 and GoogLeNet.It meets the requirements for online identification of hot-rolled steel surface defects and has promising application prospects.