计算机工程 ›› 2019, Vol. 45 ›› Issue (10): 246-252,259.doi: 10.19678/j.issn.1000-3428.0052774

• 图形图像处理 • 上一篇    下一篇

基于ACGAN的图像识别算法

周林勇1,2, 谢晓尧1, 刘志杰1, 谭宏卫2, 游善平1   

  1. 1. 贵州师范大学 信息与计算科学重点实验室, 贵阳 550001;
    2. 贵州财经大学 数学与统计学院, 贵阳 550001
  • 收稿日期:2018-09-28 修回日期:2018-11-01 出版日期:2019-10-15 发布日期:2019-10-09
  • 作者简介:周林勇(1987-),男,博士研究生,主研方向为图像处理、深度学习;谢晓尧(通信作者),教授、博士生导师;刘志杰,教授、博士;谭宏卫、游善平,博士研究生。
  • 基金项目:
    国家自然科学基金(U1631132)。

Image Identification Algorithm Based on ACGAN

ZHOU Linyong1,2, XIE Xiaoyao1, LIU Zhijie1, TAN Hongwei2, YOU Shanping1   

  1. 1. Key Laboratory of Information and Computing Science, Guizhou Normal University, Guiyang 550001, China;
    2. School of Mathematics and Statistics, Guizhou University of Finance and Economics, Guiyang 550001, China
  • Received:2018-09-28 Revised:2018-11-01 Online:2019-10-15 Published:2019-10-09

摘要: 针对基于辅助分类器生成对抗网络(ACGAN)的图像分类算法在训练过程中稳定性低且分类效果差的问题,提出一种改进的图像识别算法CP-ACGAN。对于网络结构,在判别网络的输出层取消样本的真假判别,只输出样本标签的后验估计并引入池化层。对于损失函数,除真实样本的交叉熵损失外,在判别网络中增加生成样本的条件控制标签及后验估计间的交叉熵损失。在此基础上,利用真假样本的交叉熵损失及属性重构生成器和判别器的损失函数。在MNSIT、CIFAR10、CIFAR100数据集上的实验结果表明,与ACGAN算法、CNN算法相比,该算法具有较好的分类效果与稳定性,且分类准确率分别高达99.62%、79.07%、48.03%。

关键词: 生成对抗网络, 辅助分类器生成对抗网络, 特征提取, 图像分类, 特征匹配

Abstract: To address the problem that the image classification algorithm based on Auxiliary Classifier Generative Adversarial Net(ACGAN) is unstable and the classification effect is poor,an improved image recognition algorithm CP-ACGAN is proposed.In the network structure,the authenticity discrimination of the output layer samples is cancelled.The posterior estimation of the sample label is outputted and introduced into the pooling layer.For the loss function,in addition to the cross entropy loss of real samples the cross entropy loss between the conditional control label of the generated sample and its posterior estimation is added to the discriminant network.The loss functions of the generator and discriminator are reconstructed based on the cross entropy loss and attributes of true and false samples.Experiments on MNSIT,CIFAR10 and CIFAR100 datasets show that compared with ACGAN algorithm and CNN algorithm,the algorithm has better classification effect and stability,and the classification accuracy rate is 99.62%,79.07% and 48.03% respectively.

Key words: Generative Adversarial Network(GAN), Auxiliary Classifier Generative Adversarial Network(ACGAN), feature extraction, image classification, feature matching

中图分类号: