计算机工程 ›› 2019, Vol. 45 ›› Issue (1): 17-22.doi: 10.19678/j.issn.1000-3428.0049905

• 体系结构与软件技术 • 上一篇    下一篇

面向移动平台的轻量级卷积神经网络架构

胡挺1,2,祝永新1,田犁1,封松林1,汪辉1   

  1. 1.中国科学院上海高等研究院,上海 201210; 2.中国科学院大学,北京 100049
  • 收稿日期:2017-12-28 出版日期:2019-01-15 发布日期:2019-01-15
  • 作者简介:胡挺(1992—),男,硕士研究生,主研方向为深度学习、目标检测、网络压缩;祝永新,研究员、博士;田犁,副研究员、博士;封松林、汪辉(通信作者),研究员、博士
  • 基金项目:

    国家重点研发计划(2017YFA0206104);上海市科学技术委员会科研计划项目(16511108701);上海市张江管委会公共服务平台项目(2016-14)

Lightweight Convolutional Neural Network Architecture for Mobile Platforms

HU Ting 1,2,ZHU Yongxin 1,TIAN Li 1,FENG Songlin 1,WANG Hui 1   

  1. 1.Shanghai Advanced Research Institute,Chinese Academy of Sciences,Shanghai 201210,China; 2.University of Chinese Academy of Sciences,Beijing 100049,China
  • Received:2017-12-28 Online:2019-01-15 Published:2019-01-15

摘要:

针对深度神经网络在移动平台上存在准确度低、过拟合等问题,提出一种轻量级的卷积神经网络架构。将3×3的深度可分离卷积替换SqueezeNet网络模型基本模块Fire中的标准3×3卷积核,并构建SparkNet的网络结构,替换模型卷积得到网络变形结构。实验结果表明,与SqueezeNet网络结构相比,该架构可以提高网络模型的计算速度,有效降低网络模型规模并减少参数数量。

关键词: 深度学习, 卷积神经网络, 深度可分离卷积, 神经网络压缩, 轻量级

Abstract:

For the problem that the deep neural network has low accuracy and over-fitting on the mobile platforms,a lightweight Convolutional Neural Network(CNN) architecture is proposed.The 3×3 depthwise separable convolution replaces the standard 3×3 convolution kernel in the SqueezeNet network model basic module Fire,constructs the SparkNet network structure,and replaces the model convolution to obtain the network deformation structure.Experimental results show that compared with the SqueezeNet network structure,the architecture can improve the calculation speed of the network model,effectively reduce the network model size and reduce the number of parameters.

Key words: deep learning, Convolutional Neural Network(CNN), depthwise separable convolution, neural network compression, lightweight

中图分类号: