基于超轻子空间注意力的深度学习电子元件分类  

Classification of electronic components using deep learning based on ultra⁃lightweight subspace attention

作　　者：杜慧江 徐波 秦梦雨 孙丽萍 DU Huijiang;XU Bo;QIN Mengyu;SUN Liping(School of Medical Instruments,Shanghai University of Medicine and Health Sciences,Shanghai 201318,China)

机构地区：[1]上海健康医学院医疗器械学院,上海201318

出　　处：《电子设计工程》2025年第5期13-17,共5页Electronic Design Engineering

基　　金：国家重点研发计划资助项目(2018YFB1307700)。

摘　　要：基于提高在电子产品的生产和回收领域对电子元件分类的速度和准确率的目的,提出一种基于超轻子空间注意力的电子元件分类深度学习模型。在使用4个卷积层的基线模型上嵌入超轻子空间注意力模块,在公开的电子元件数据集的包含14个类别的子集上调优选择超参数进行训练。实验结果显示,该模型在公开电子元件数据集上的分类准确率达到95.09%,高于基线模型和用于对比的不加载ImageNet权重的ResNet50V2、EfficientNetV2S、MobileNetV3Large模型,并且参数量更小、训练速度更快,可以有效提升电子元件分类的准确率和速度。To enhance the speed and accuracy of classifying electronic components in both production and recycling settings,a deep learning model equipped with an ultra-lightweight subspace attention module is introduced.The module is integrated into a baseline architecture comprising four convolutional layers.Hyperparameter tuning and rigorous training are conducted on a subset of a publicly accessible dataset containing 14 distinct categories of electronic components.Our experimental findings reveal that the proposed model attains a remarkable classification accuracy of 95.09%on the public benchmark,outperforming the baseline as well as several state-of-the-art models-namely,ResNet50V2,Efficient-NetV2S and MobileNetV3Large—all evaluated without ImageNet pre-training.Notably,our approach boasts reduced parameter counts,and swifter training times,collectively contributing to significant improvements in both the precision and efficiency of electronic component classification.

关 键 词：轻量 空间注意力 深度学习 电子元件 分类 

分 类 号：TN6[电子电信—电路与系统]