基于CBAM-ResNet50的民国纸币图像检索系统  

Image retrieval system of the Republic of China banknotes based on CBAM-ResNet50

作　　者：王佳婧 朱媛媛[1] 杜欣 王笑梅[1] WANG Jiajing;ZHU Yuanyuan;DU Xin;WANG Xiaomei(College of Information,Mechanical and Electrical Engineering,Shanghai Normal University,Shanghai 201418,China)

机构地区：[1]上海师范大学信息与机电工程学院,上海201418

出　　处：《上海师范大学学报（自然科学版）》2022年第4期414-419,共6页Journal of Shanghai Normal University(Natural Sciences)

摘　　要：利用残差网络(ResNet)50,结合卷积块注意力模块(CBAM)机制,提出了一种基于CBAM-ResNet50的民国纸币图像检索技术,提升了对相似纸币的检索能力.设计并实现了基于Windows和Ubuntu系统环境下的民国纸币图像检索系统,并搭建了基于Flask的Web应用服务.所提取的民国纸币图像特征具有更强的辨识度,大幅提高了检索速度,在图形处理器(GPU)上可达毫秒级.使用缩略图搜索民国纸币图片,对相似度排名第1的图像的检索准确率可以达76.3%,相似度排名前6的图像检索准确率可以达92.5%.Using the residual network(ResNet)50 backbone network,and combining convolutional block attention module(CBAM)mechanism,we propose a kind of technique for image retrieval of banknotes based on CBAM-ResNet50.We improve the retrieval of similar banknotes.After building a Flask-based web application service,an image retrieval system for banknotes based on Windows and Ubuntu system environment is designed and implemented.The extracted image features of banknotes are more recognizable,and the retrieval speed is improved greatly,which can reach millisecond-level on graphics processing unit(GPU).Using thumbnails to search banknotes image,the image accuracy with the first similarity ranking can reach 76.3%.The image accuracy of the top 6 in similarity ranking can reach 92.5%.

关 键 词：深度学习 残差网络(ResNet) 民国纸币 图像检索 卷积块注意力模块(CBAM) 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术] TP183[自动化与计算机技术—计算机科学与技术]