计算机工程

• 人工智能及识别技术 • 上一篇    下一篇

分离多路卷积神经网络研究

宋超,许道云,秦永彬   

  1. (贵州大学 计算机科学与技术学院,贵阳 550025)
  • 收稿日期:2016-08-04 出版日期:2017-06-15 发布日期:2017-06-15
  • 作者简介:宋超(1989—),男,硕士研究生,主研方向为图像识别;许道云(通信作者),教授、博士生导师;秦永彬,副教授、博士。
  • 基金项目:
    国家自然科学基金(61262006,61540050);贵州省重大应用基础研究项目(黔科合JZ字[2014]2001);贵州省科技厅联合基金(黔科合LH字[2014]7636号)。

Research on Detached Multiple Convolutional Neural Network

SONG Chao,XU Daoyun,QIN Yongbin   

  1. (College of Computer Science and Technology,Guizhou University,Guiyang 550025,China)
  • Received:2016-08-04 Online:2017-06-15 Published:2017-06-15

摘要: 针对卷积神经网络主要使用图像的局部特征而忽略图像通道特征的不足,提出一种分离多路卷积神经网络。提取通道特征与卷积特征,并在全连接层进行融合,以此提升该网络的图像识别与分类效果。在cifar10和SVHN数据集上进行的实验结果表明,与ResNet,Network in Network,Maxout等8种卷积神经网络相比,该网络的平均识别率较高。

关键词: 卷积神经网络, 深度学习, 特征提取, 图像分类, 图像识别, 通道特征

Abstract: As the Convolutional Neural Network(CNN) mainly uses the local features of the image,ignoring image channel features,this paper proposes a Detached Multiple Convolutional Neural Network(DMCNN).It extracts the channel features and convolution features,and fuses them in the whole connection layer so that image recognition and classification effects of the proposed network are improved.The experimental results on cifar10 and SVHN datasets show that the average recognition rate of the network is higher than that of other 8 CNNs like RexNet,Network in Network, Maxout.

Key words: Convolutional Neural Network(CNN), deep learning, feature extraction, image classification, image recognition, channel characteristic

中图分类号: