Journal of Computer Applications ›› 2018, Vol. 38 ›› Issue (11): 3305-3311.DOI: 10.11772/j.issn.1001-9081.2018051008

Previous Articles     Next Articles

Data augmentation method based on conditional generative adversarial net model

CHEN Wenbing, GUAN Zhengxiong, CHEN Yunjie   

  1. School of Mathematics and Statistics, Nanjing University of Information Science & Technology, Nanjing Jiangsu 210044, China
  • Received:2018-05-14 Revised:2018-06-26 Online:2018-11-10 Published:2018-11-10
  • Supported by:
    This work is partially supported by the National Natural Science Foundation of China (61672291), the Beijige Foundation (BJG201504).

基于条件生成式对抗网络的数据增强方法

陈文兵, 管正雄, 陈允杰   

  1. 南京信息工程大学 数学与统计学院, 南京 210044
  • 通讯作者: 管正雄
  • 作者简介:陈文兵(1964-),男,安徽东至人,副教授,硕士,主要研究方向:计算数学、模式识别、图像处理;管正雄(1993-),男,安徽芜湖人,硕士研究生,主要研究方向:模式识别、图像处理;陈允杰(1980-),男,江苏南京人,教授,博士,主要研究方向:计算数学、模式识别、图像处理。
  • 基金资助:
    国家自然科学基金资助项目(61672291);北极阁基金资助项目(BJG201504)。

Abstract: Deep Convolutional Neural Network (CNN) is trained by large-scale labelled datasets. After training, the model can achieve high recognition rate or good classification effect. However, the training of CNN models with smaller-scale datasets usually occurs overfitting. In order to solve this problem, a novel data augmentation method called GMM-CGAN was proposed, which was integrated Gaussian Mixture Model (GMM) and CGAN (Conditional Generative Adversarial Net). Firstly, sample number was increased by randomly sliding sampling around the core region. Secondly, the random noise vector was supposed to submit to the distribution of GMM model, then it was used as the initial input to the CGAN generator and the image label was used as the CGAN condition to train the parameters of the CGAN and GMM models. Finally, the trained CGAN was used to generate a new dataset that matched the real distribution of the samples. The dataset was divided into 12 classes of 386 items. After implementing GMM-CGAN on the dataset, the total number of the new dataset was 38600. The experimental results show that compared with CNN's training datasets augmented by Affine transformation or CGAN, the average classification accuracy of the proposed method is 89.1%, which is improved by 18.2% and 14.1%, respectively.

Key words: image classification, deep Convolution Neural Network (CNN), Gaussian Mixture Model (GMM), Conditional Generative Adversarial Net (CGAN), data augmentation algorithm

摘要: 深度卷积神经网络(CNN)在大规模带有标签的数据集训练下,训练后模型能够取得高的识别率或好的分类效果,而利用较小规模数据集训练CNN模型则通常出现过拟合现象。针对这一问题,提出了一种集成高斯混合模型(GMM)及条件生成式对抗网络(CGAN)的数据增强方法并记作GMM-CGAN。首先,通过围绕核心区域随机滑动采样的方法增加数据集样本数量;其次,假定噪声随机向量服从GMM描述的分布,将它作为CGAN生成器的初始输入,图像标签作为CGAN条件,训练CGAN以及GMM模型的参数;最后,利用已训练CGAN生成符合样本真实分布的新数据集。对包含12种雾型386个样本的天气形势图基准集利用GMM-CGAN方法进行数据增强,增强后的数据集样本数多达38600个,将该数据集训练的CNN模型与仅使用仿射变换增强的数据集及CGAN方法增强的数据集训练的CNN模型相比,实验结果表明,前者的平均分类正确率相较于后两个模型分别提高了18.2%及14.1%,达到89.1%。

关键词: 图像分类, 深度卷积神经网络, 高斯混合模型, 有条件对抗神经网络, 数据增强算法

CLC Number: