融合CBAM机制的ResNet34模型用于电子基板玻璃热工缺陷分类研究

doi:10.16552/j.cnki.issn1001-1625.2025.1071

硅酸盐通报 ›› 2026, Vol. 45 ›› Issue (3): 1074-1082.DOI: 10.16552/j.cnki.issn1001-1625.2025.1071

融合CBAM机制的ResNet34模型用于电子基板玻璃热工缺陷分类研究

曹志强¹^,²(), 李苑², 金良茂³, 于浩³, 曹欣³, 刘涌²(), 韩高荣²^,⁴

^1.浙江大学先进玻璃材料全国重点实验室，杭州 310058
^2.浙江大学材料科学与工程学院，杭州 310058
^3.中建材玻璃新材料研究院集团有限公司先进玻璃材料全国重点实验室，蚌埠 233010
^4.浙江大学宁波国际科创中心，宁波 315100

收稿日期:2025-10-31 修订日期:2025-11-30 出版日期:2026-03-20 发布日期:2026-04-10
通信作者: 刘涌，博士，副教授。E-mail：liuyong.mse@zju.edu.cn
作者简介:曹志强（1983—），男，博士研究生。主要从事电子玻璃的研究。E-mail：caozq@ctiec.net
基金资助:
国家“十四五”重点研发计划(2022YFB3603300)

Classification of Thermal Defects in Electronic Substrate Glass with ResNet34 Model Integrating CBAM Mechanism

CAO Zhiqiang¹^,²(), LI Yuan², JIN Liangmao³, YU Hao³, CAO Xin³, LIU Yong²(), HAN Gaorong²^,⁴

^1.State Key Laboratory of Advanced Glass Materials，Zhejiang University，Hangzhou 310058，China
^2.School of Materials Science and Engineering，Zhejiang University，Hangzhou 310058，China
^3.State Key Laboratory of Advanced Glass Materials，CNBM Research Institute for Advanced Glass;Materials Group Co. ，Ltd. ，Bengbu 233010，China
^4.Ningbo Global Innovation Center，Zhejiang University，Ningbo 315100，China

Received:2025-10-31 Revised:2025-11-30 Published:2026-03-20 Online:2026-04-10

摘要/Abstract

摘要：

电子基板玻璃是信息显示产业的关键基础材料之一。近年来信息显示产业向大尺寸、超高清和轻薄化发展，对电子基板玻璃的质量提出了更高的要求。本文针对电子基板玻璃热工缺陷尺寸小、相似度高、识别难度大的问题，以深度残差网络模型（ResNet34）为主体框架，引入卷积块注意力模块（CBAM）增强对小目标缺陷的感知能力。结果表明，基于自建的六类典型热工缺陷数据集，融合CBAM机制的ResNet34模型的分类准确率从95.74%增至98.08%，同时泛化能力得到明显提升。可视化分析比较表明，CBAM机制对缺陷识别能力的提升来自对缺陷的准确定位。以上结果为电子基板玻璃热工缺陷的在线智能检测提出了一种可行方案，也可作为其他小目标分类的参考。

关键词: 电子基板玻璃, 玻璃熔制, 深度卷积神经网络, 注意力机制, 热工缺陷, 缺陷分类

Abstract:

Electronic substrate glass is one of the key fundamental material in the information display industry. In recent years， as the information display industry has moved toward larger sizes， ultra-high definitions， and thinner， lighter designs， higher standards have been set for the quality of electronic substrate glass. In this paper， aiming at the challenges such as small size， high similarity， and difficult recognition of thermal defects in electronic substrate glass， the deep residual network model （ResNet34） was used as the main framework， the convolutional block attention module （CBAM） was integrated to improve the detection of minor target defects. The results show that， based on a homemade dataset of six typical thermal defects， the classification accuracy of the ResNet34 model integrating CBAM mechanism increases from 95.74% to 98.08%， with a notable improvement in generalization. Further visual analysis and comparisons reveal that the improved defect recognition performance with CBAM results from more precise defect localization. These findings offer a practical solution for online intelligent detection of thermal defects in electronic substrate glass and may serve as a reference for other minor target classification.

Key words: electronic substrate glass, glass manufacture, deep convolutional neural network, attention mechanism, thermal defect, defect classification

中图分类号:

TQ171

曹志强, 李苑, 金良茂, 于浩, 曹欣, 刘涌, 韩高荣. 融合CBAM机制的ResNet34模型用于电子基板玻璃热工缺陷分类研究[J]. 硅酸盐通报, 2026, 45(3): 1074-1082.

CAO Zhiqiang, LI Yuan, JIN Liangmao, YU Hao, CAO Xin, LIU Yong, HAN Gaorong. Classification of Thermal Defects in Electronic Substrate Glass with ResNet34 Model Integrating CBAM Mechanism[J]. BULLETIN OF THE CHINESE CERAMIC SOCIETY, 2026, 45(3): 1074-1082.

图/表 8

图1 典型玻璃热工缺陷图

Fig.1 Typical glass thermal defect images

表1 深度残差网络模型ResNet34的配置

Table 1 Configuration for deep residual network model ResNet34

Layer	Convolutional kernel	Number of channel	Number of stack	Output size of feature map/Pixel
Input	—	3	—	3×384×64
Pre-conv	$7 × 7$	64	1	64×192×32
Maxpool	$3 × 3$	64	1	64×192×32
ResNet-conv1	$3 × 3 3 × 3$	$6464$	3	64×96×16
ResNet-conv2	$3 × 3 3 × 3$	$128128$	4	128×48×8
ResNet-conv3	$3 × 3 3 × 3$	$256256$	6	256×24×4
ResNet-conv4	$3 × 3 3 × 3$	$512512$	3	512×12×2
Output	Avg-pool+flatten+fully connected+softmax			Class

表1 深度残差网络模型ResNet34的配置

Table 1 Configuration for deep residual network model ResNet34

Layer	Convolutional kernel	Number of channel	Number of stack	Output size of feature map/Pixel
Input	—	3	—	3×384×64
Pre-conv	$7 × 7$	64	1	64×192×32
Maxpool	$3 × 3$	64	1	64×192×32
ResNet-conv1	$3 × 3 3 × 3$	$6464$	3	64×96×16
ResNet-conv2	$3 × 3 3 × 3$	$128128$	4	128×48×8
ResNet-conv3	$3 × 3 3 × 3$	$256256$	6	256×24×4
ResNet-conv4	$3 × 3 3 × 3$	$512512$	3	512×12×2
Output	Avg-pool+flatten+fully connected+softmax			Class

图2 融合CBAM机制的ResNet34结构示意图

Fig.2 Schematic diagram of ResNet34 structure integrating CBAM mechanism

图3 模型训练过程准确率曲线

Fig.3 Accuracy curves during model training process

表2 消融试验的模型评价指标

Table 2 Model evaluation indicators for ablation experiments

Group	Test set	Training set				Model size/MB	Total training time/（h∶min∶s）
Group	Acc/%	Acc/%	Pr/%	Re/%	F1/%	Model size/MB	Total training time/（h∶min∶s）
A	85.80	95.74	94.60	94.58	94.59	123.46	1∶10∶12
B	88.41	96.50	96.70	96.67	96.68	129.03	1∶32∶46
C	91.84	97.51	97.97	97.92	97.94	141.64	1∶35∶40
D	93.56	98.08	98.35	98.33	98.34	141.64	1∶46∶05

图4 不同模型在数据集上分类结果的混淆矩阵

Fig.4 Confusion matrix of classification results of different models on dataset

图5 不同模型沾锡缺陷的梯度类激活映射图

Fig.5 Gradient class activation mapping of tin defects in different models

图6 不同模型同一缺陷的梯度类激活映射图

Fig.6 Gradient class activation mapping of same defect in different models

参考文献 20

[1]	王红军. 机器视觉——现代工业的眼睛［J］. 机电一体化， 1999， 5（3）： 26-27.
	WANG H J. Machine vision： eyes of modern industry［J］. Mechatronics， 1999， 5（3）： 26-27 （in Chinese）.
[2]	明五一，贾豪杰，何文斌，等. 透明件表面缺陷的机器视觉检测综述［J］. 机械科学与技术， 2021， 40（1）： 116-124.
	MING W Y， JIA H J， HE W B， et al. Detecting surface of transparent parts with computer vision［J］. Mechanical Science and Technology for Aerospace Engineering， 2021， 40（1）： 116-124.
[3]	LIN S Y， WU Y Q. Vision-based LCD/OLED defect detection methods： a critical summary［J］. Journal of Image and Graphics， 2024， 29（5）： 1321-1345.
[4]	LEE S， KIM B， LEE H. Deep learning-based glass detection for smart glass manufacturing processes［J］. Computers， Materials & Continua， 2025， 84（1）： 1397-1415.
[5]	YE S J， WANG Z， XIONG P B， et al. Multi-stage few-shot micro-defect detection of patterned OLED panel using defect inpainting and multi-scale Siamese neural network［J］. Journal of Intelligent Manufacturing， 2024， 35（6）： 2653-2669.
[6]	LI L L， WANG Z F， ZHANG T T. GBH-YOLOv5： ghost convolution with BottleneckCSP and tiny target prediction head incorporating YOLOv5 for PV panel defect detection［J］. Electronics， 2023， 12（3）： 561.
[7]	JHA S B， BABICEANU R F. Deep CNN-based visual defect detection： survey of current literature［J］. Computers in Industry， 2023， 148： 103911.
[8]	LIU H， SUN X， LI Y， et al. Review of deep learning models for Image classification based on convolutional neural networks［J］. Computer Engineering and Application， 2025， 61（11）： 1-21.
[9]	KRIZHEVSKY A， SUTSKEVER I， HINTON G E. ImageNet classification with deep convolutional neural networks［J］. Communications of the ACM， 2017， 60（6）： 84-90.
[10]	SIMONYAN K， ZISSERMAN A. Very deep convolutional networks for large-scale image recognition［C］//3rd International Conference on Learning Representations （ICLR 2015）， Computational and Biological Learning Society， 2015： 1-14.
[11]	HE K M， ZHANG X Y， REN S Q， et al. Deep residual learning for image recognition［C］//2016 IEEE Conference on Computer Vision and Pattern Recognition （CVPR）， 2016： 770-778.
[12]	HU J， SHEN L， ALBANIE S， et al. Squeeze-and-excitation networks［J］. IEEE Transactions on Pattern Analysis and Machine Intelligence， 2019， 42（8）： 2011-2023.
[13]	WOO S H， PARK J， LEE J Y， et al. CBAM： convolutional block attention module［C］//Computer Vision-ECCV 2018： 15th European Conference， 2018： 3-19.
[14]	刘红玉，高见. 融合CBAM的违法犯罪类安卓恶意软件检测与分类模型研究［J］计算机工程与应用， 2025， 61（6）： 317-327.
	LIU H Y， GAO J. Research on detection and classification model of illegal and criminal android malware integrating CBAM［J］. Computer Engineering and Applications， 2025， 61（6）： 317-327 （in Chinese）.
[15]	QI J D， WANGDUI B B， JIANG J， et al. EDKSANet： an efficient dual-kernel split attention neural network for the classification of Tibetan medicinal materials［J］. Electronics， 2023， 12（20）： 4330.
[16]	LIN G J， LIU K Y， XIA X K， et al. An efficient and intelligent detection method for fabric defects based on improved YOLOv5［J］. Sensors， 2023， 23（1）： 97.
[17]	LIU B， WANG T T， LIU J. Surface defect detection of mobile phone covers based on improved BiSeNet V2［J］. Acta Optica Sinica， 2024， 44（16）： 1615001.
[18]	欧阳健燊. 基于深度学习的手机屏幕缺陷检测研究［D］. 广州：华南理工大学， 2020.
	OUYANG J S. Research on mobile phone screen defect detection method based on deep learning［D］. Guangzhou： South China University of Technology， 2020 （in Chinese）.
[19]	PARK J， RIAZ H， KIM H， et al. Advanced cover glass defect detection and classification based on multi-DNN model［J］. Manufacturing Letters， 2020， 23： 53-61.
[20]	SELVARAJU R R， COGSWELL M， DAS A， et al. Grad-CAM： visual explanations from deep networks via gradient-based localization［C］//2017 IEEE International Conference on Computer Vision （ICCV）， 2017： 618-626.

融合CBAM机制的ResNet34模型用于电子基板玻璃热工缺陷分类研究

Classification of Thermal Defects in Electronic Substrate Glass with ResNet34 Model Integrating CBAM Mechanism

RichHTML

PDF (PC)

可视化

摘要/Abstract

引用本文

使用本文

图/表 8

参考文献 20

相关文章 2

编辑推荐

Metrics

本文评价

[1]	刘晨宇, 张文露, 李从云, 曾健华, 李路瑶, 韩建军, 王静. 低碳玻璃配方对玻璃熔制过程及CO₂排放的影响[J]. 硅酸盐通报, 2026, 45(3): 1083-1093.
[2]	朱经纬, 舒众众, 金良茂, 曹志强, 张冲, 郑际杰, 刘涌, 韩高荣. 残余应力分布对基板玻璃落球冲击强度影响的数值模拟研究[J]. 硅酸盐通报, 2024, 43(4): 1267-1273.