计算机工程与应用 ›› 2017, Vol. 53 ›› Issue (13): 181-188.DOI: 10.3778/j.issn.1002-8331.1601-0435

• 图形图像处理 • 上一篇    下一篇

基于深度卷积神经网络的快速图像分类算法

王华利1,邹俊忠1,张  见1,卫作臣1,汪春梅2   

  1. 1.华东理工大学 信息科学与工程学院,上海 200237
    2.上海师范大学 信息与机电工程学院,上海 200234
  • 出版日期:2017-07-01 发布日期:2017-07-12

Fast image classification algorithm based on deep convolutional neural network

WANG Huali1, ZOU Junzhong1, ZHANG Jian1, WEI Zuochen1, WANG Chunmei2   

  1. 1.School of Information Science and Engineering, East China University of Science and Technology, Shanghai 200237, China
    2.School of Information, Mechanical and Electrical Engineering, Shanghai Normal University, Shanghai 200234, China
  • Online:2017-07-01 Published:2017-07-12

摘要: 为了应对大量图像的分类问题,提出一种基于深度卷积神经网络和CUDA-cuDNN并行运算的快速图像分类方法。该方法利用深度卷积神经网络自动学习特征的优势来解决手工设计特征普适性差等问题,同时结合基于CUDA架构的cuDNN并行运算策略来提高训练速度和加快分类速度,并且针对深度卷积神经网络易受参数扰动等缺点,引入批量正则化(Batch Normalization)以提高算法的鲁棒性。实验结果表明,该方法不仅大幅缩短了训练时间同时加快了图像的分类速度,而且进一步降低了图像分类的错误率。

关键词: 深度卷积神经网络, CUDA-cuDNN方法, 批量正则化, 图像分类, 深度学习

Abstract: In order to solve large amount of images classification issues, a method is introduced by combining with CUDA-cuDNN and Deep Convolutional Neural Network(DCNN). This method makes advantages of DCNNs to learn features automatically, which makes up the incapability of hand-crafted features. Meanwhile, a cuDNN parallel computing method based on CUDA is employed to improve the speed of training and validation. DCNN is susceptible to parameter perturbation, which employs Batch Normalization(BN) to enhance the robustness. Experimental results indicate that the proposed method not only reduces training time substantially and accelerates validation speed, but also obtains lower classification error rate.

Key words: Deep Convolutional Neural Network(DCNN), CUDA-cuDNN, batch normalization, image classification, deep learning