作者投稿和查稿 主编审稿 专家审稿 编委审稿 远程编辑

计算机工程 ›› 2021, Vol. 47 ›› Issue (1): 305-311. doi: 10.19678/j.issn.1000-3428.0056239

• 开发研究与工程应用 • 上一篇    下一篇

基于注意力机制的商家招牌分类研究

李兰1, 郑雨薇1, 魏少玮2, 胡克勇1   

  1. 1. 青岛理工大学 信息与控制工程学院, 山东 青岛 266520;
    2. 西安电子科技大学 人工智能学院, 西安 710071
  • 收稿日期:2019-10-10 修回日期:2020-01-03 发布日期:2020-04-10
  • 作者简介:李兰(1963-),女,教授,主研方向为数据挖掘、模式识别;郑雨薇(通信作者)、魏少玮,硕士研究生;胡克勇,副教授、博士。
  • 基金资助:
    国家自然科学基金“空气质量时空联合重构关键技术研究”(61902205)。

Research on Classification of Business Signs Based on Attention Mechanism

LI Lan1, ZHENG Yuwei1, WEI Shaowei2, HU Keyong1   

  1. 1. School of Information and Control Engineering, Qingdao University of Technology, Qingdao, Shandong 266520, China;
    2. School of Artificial Intelligence, Xidian University, Xi'an 710071, China
  • Received:2019-10-10 Revised:2020-01-03 Published:2020-04-10

摘要: 为解决采用卷积神经网络对商家招牌进行分类时存在特征判别性较差的问题,通过在注意力机制中引入神经网络,提出一种端到端的深度学习卷积神经网络方法。使用卷积注意力模块分别学习通道注意力与空间注意力信息以增强特征的判别性,利用余弦间隔损失函数增强所提取特征的泛化能力,且可在特征空间中减小类内方差与增大类间间隔。实验结果表明,与基于传统交叉损失函数方法相比,该方法通过将注意力机制模块与余弦间隔损失函数相结合,使得准确率与F1值分别提高2.2和2.0个百分点,达到99.3%和98.6%。

关键词: 端到端的深度学习, 卷积神经网络, 注意力机制, 余弦间隔损失函数, 商家招牌分类

Abstract: In order to solve the problem of poor feature discrimination when using convolutional neural network to classify business signs,an end-to-end deep learning convolutional neural network method is proposed by introducing the neural network into the attention mechanism.The convolutional attention module is used to learn channel attention and spatial attention information respectively to enhance the discriminability of features.The cosine interval loss function is used to enhance the generalization ability of feature extraction,and to reduce the intra-class variance and increase the inter-class interval in the feature space.The experimental results show that compared with the method based on the traditional cross loss function,the proposed method improves the accuracy by 2.2 percentage points to 99.3%,and F1 value by 2.0 percentage points to 98.6% after combining the attention mechanism module with the cosine interval loss function.

Key words: end-to-end deep learning, Convolutional Neural Network(CNN), attention mechanism, cosine interval loss function, classification of business signs

中图分类号: