《计算机应用》唯一官方网站 ›› 2022, Vol. 42 ›› Issue (8): 2353-2360.DOI: 10.11772/j.issn.1001-9081.2021061037

• 人工智能 • 上一篇    下一篇

基于挤压激励的轻量化注意力机制模块

吕振虎1, 许新征1,2(), 张芳艳3   

  1. 1.中国矿业大学 计算机科学与技术学院, 江苏 徐州 221116
    2.光电技术与智能控制教育部重点实验室(兰州交通大学), 兰州 730070
    3.宁夏大学 智能工程与技术学院, 宁夏 中卫 755000
  • 收稿日期:2021-06-21 修回日期:2021-09-04 接受日期:2021-09-14 发布日期:2021-10-18 出版日期:2022-08-10
  • 通讯作者: 许新征
  • 作者简介:吕振虎(1995—),男,山东枣庄人,硕士研究生,主要研究方向:机器学习、计算机视觉;
    许新征(1980—),男,安徽宿州人,教授,博士,CCF高级会员,主要研究方向:机器学习、模式识别;
    张芳艳(1990—),女,甘肃宁县人,硕士,主要研究方向:图像处理、模式识别。
  • 基金资助:
    国家自然科学基金资助项目(61976217);光电技术与智能控制教育部重点实验室(兰州交通大学)开放课题(KFKT2020-03)

Lightweight attention mechanism module based on squeeze and excitation

Zhenhu LYU1, Xinzheng XU1,2(), Fangyan ZHANG3   

  1. 1.School of Computer Science and Technology,China University of Mining and Technology,Xuzhou Jiangsu 221116,China
    2.Key Laboratory of Opt-Electronic Technology and Intelligent Control of Ministry of Education (Lanzhou Jiaotong University),Lanzhou Gansu 730070,China
    3.School of Intelligent Engineering and Technology,Ningxia University,Zhongwei Ningxia 755000,China
  • Received:2021-06-21 Revised:2021-09-04 Accepted:2021-09-14 Online:2021-10-18 Published:2022-08-10
  • Contact: Xinzheng XU
  • About author:LYU Zhenhu, born in 1995, M. S. candidate. His research interests include machine learning, computer vision.
    XU Xinzheng, born in 1980, Ph. D., professor. His research interests include machine learning, pattern recognition.
    ZHANG Fangyan, born in 1990, M. S. Her research interests include image processing, pattern recognition.
  • Supported by:
    National Natural Science Foundation of China(61976217);Opening Project of Key Laboratory of Opt-Electronic Technology and Intelligent Control of Ministry of Education (Lanzhou Jiaotong University)(KFKT2020-03)

摘要:

针对向卷积神经网络(CNN)中嵌入注意力机制模块以提高模型应用精度导致参数和计算量增加的问题,提出基于挤压激励的轻量化高度维度挤压激励(HD-SE)模块和宽度维度挤压激励(WD-SE)模块。为了充分利用特征图中潜在的信息,HD-SE对卷积层输出的特征图在高度维度上进行挤压激励操作,获得高度维度上的权重信息;而WD-SE在宽度维度上进行挤压激励操作,以得到特征图宽度维度上的权重信息;然后,将得到的权重信息分别应用于对应维度的特征图张量,以提高模型的应用精度。将HD-SE与WD-SE分别嵌入VGG16、ResNet56、MobileNetV1和MobileNetV2模型中,在CIFAR10和CIFAR100数据集上进行的实验结果表明,与挤压激励(SE)模块、协调注意力(CA)模块、卷积块注意力模块(CBAM)和高效通道注意力(ECA)模块等先进的注意力机制模块相比,HD-SE与WD-SE在向网络模型中增加的参数和计算量更少的同时得到的精度相似或者更高。

关键词: 卷积神经网络, 挤压激励, 轻量化, 多维度, 注意力机制模块

Abstract:

Focusing on the issue that embedding the attention mechanism module into Convolutional Neural Network (CNN) to improve the application accuracy will increase the parameters and the computational cost, the lightweight Height Dimensional Squeeze and Excitation (HD-SE) module and Width Dimensional Squeeze and Excitation (WD-SE) module based on squeeze and excitation were proposed. To make full use of the potential information in the feature maps, two kinds of height and width dimensional weight information of feature maps was respectively extracted by HD-SE and WD-SE through squeeze and excitation operations, then the obtained weight information was respectively applied to corresponding tensors of the feature maps of two dimensions to improve the application accuracy of the model. Experiments were implemented on CIFAR10 and CIFAR100 datasets after embedding HD-SE and WD-SE into Visual Geometry Group 16 (VGG16), Residual Network 56 (ResNet56), MobileNetV1 and MobileNetV2 models respectively. Experimental results show fewer parameters and computational cost added by HD-SE and WD-SE to the network models when the models achieve the same or even better accuracy, compared with the state-of-the-art attention mechanism modules, such as Squeeze and Excitation (SE) module, Coordinate Attention (CA) block, Convolutional Block Attention Module (CBAM) and Efficient Channel Attention (ECA) module.

Key words: Convolutional Neural Network (CNN), squeeze and excitation, lightweight, multi-dimension, attention mechanism module

中图分类号: