基于卷积神经网络和混合注意力机制的书标检测算法  

Title label detection using convolutional neural network and hybrid attention mechanism

作　　者：张岩[1,2] 赵蒙蒙 孙英伟 常艳康 ZHANG Yan;ZHAO Mengmeng;SUN Yingwei;CHANG Yankang(Library of Qingdao University of Science and Technology,Qingdao 266061,China;College of Electromechanical Engineering,Qingdao University of Science and Technology,Qingdao 266061,China)

机构地区：[1]青岛科技大学图书馆,山东青岛266061 [2]青岛科技大学机电工程学院,山东青岛266061

出　　处：《山东科技大学学报（自然科学版）》2023年第3期94-102,共9页Journal of Shandong University of Science and Technology(Natural Science)

基　　金：国家自然科学基金项目(62172248);山东省自然科学基金项目(ZR2019MEE066)。

摘　　要：为实现图书馆中机器人智能排架,提出一种基于卷积神经网络和混合注意力机制的书标检测模型。将DenseNet121引入YOLOv4以提高特征和梯度之间的传递效率,利用SPDC模块实现局部和全局特征融合,进而通过通道和空间混合注意力提高模型的特征表征能力。实验结果表明,模型的平均准确率、整体性能、参数量和模型大小均优于对比方法,且易于部署到嵌入式设备中实现在线检测,从而提高图书乱架治理的智能化水平。To realize robot intelligent shelving in library,this paper proposed a title label detection model based on convolutional neural network and hybrid attention mechanism.DenseNet121 was applied to YOLOv4 to improve the transfer efficiency between features and gradients.The spatial pyramid dilated convolution(SPDC)module was used to achieve local and global feature fusion.Then the model’s feature representation ability was improved through channel and spatial attention.Experimental results show that the average accuracy,overall performance,parameter amount and model size of the proposed model outperforms those of the compared methods and it is easy to deploy to embedded devices to achieve online detection,thus improving the intelligent level of book disorder management.

关 键 词：卷积神经网络 混合注意力机制 书标 目标检测 智慧图书馆 

分 类 号：TP393[自动化与计算机技术—计算机应用技术] G250.7[自动化与计算机技术—计算机科学与技术]