基于注意力机制和深度残差网络的烟盒规格识别  

CIGARETTE RECOGNITION BASED ON ATTENTION MECHANISMAND DEEP RESIDUAL NETWORK

作　　者：赵志成 罗泽[1] Zhao Zhicheng;Luo Ze(Computer Network Information Center,Chinese Academy of Sciences,Beijing 100190,China;University of Chinese Academy of Sciences,Beijing 100049,China)

机构地区：[1]中国科学院计算机网络信息中心,北京100190 [2]中国科学院大学,北京100049

出　　处：《计算机应用与软件》2023年第9期242-247,252,共7页Computer Applications and Software

基　　金：中国烟草总公司科技重大专项(110201901026(SJ-05))。

摘　　要：为了解决在人工识别烟盒规格过程中存在的识别效率低、识别错误率高、不便于批量部署等问题,基于深度残差网络和注意力机制建立烟盒规格自动识别模型。烟盒规格识别可以归纳为细粒度识别任务,单独使用深度残差网络进行细粒度分类存在特征判别性较差、精度不足的问题。自适应选择卷积网络(Selective Kernel Networks, SKNet)可以动态地调节感受野从而更好地提取图像的细节特征,通过将自适应选择卷积网络和深度残差网络相结合,不仅增强了深度残差网络提取的特征的判别性而且可以有效地提升分类的精度。与人工识别方法相比,该方法不仅实现了自动化的烟盒规格识别,而且识别的准确率达到了99.2%。In order to solve the problems such as low recognition efficiency,high recognition error rate and inconvenient batch deployment in the process of manual recognition of cigarette packet specifications,this paper establishes an automatic recognition model of cigarette packet specifications based on deep residual network and attention mechanism.Cigarette packet recognition can be summarized as a fine-grained recognition task.The use of deep residual networks alone for fine-grained classification suffers from poor feature discriminability and insufficient accuracy.Selective kernel networks(SKNet)can dynamically adjust the receptive field to better extract the detailed features of the image.The combination of deep residual networks and SKNet can not only enhance the discriminative power of the features but also effectively improve the classification of the accuracy.Compared with the manual recognition method,the proposed method not only realizes automatic identification of cigarette pack specifications,but also has an accuracy of 99.2%.

关 键 词：烟盒规格识别 卷积神经网络 深度学习 深度残差网络 注意力机制 

分 类 号：TP3[自动化与计算机技术—计算机科学与技术]