基于DCGAN的手写体蒙文字元样本增强  

Enhancement of Handwritten Mongolian Meta Samples Based on Deep Convolutional Generative Adversarial Networks

作　　者：刁明皓 戚桂美 殷雁君 DIAO Minghao;QI Guimei;YIN Yanjun(College of Computer Science and Technology,Inner Mongolia Normal University,Hohhot 010022,China)

机构地区：[1]内蒙古师范大学计算机科学技术学院,呼和浩特010022

出　　处：《内蒙古农业大学学报（自然科学版）》2021年第6期93-98,共6页Journal of Inner Mongolia Agricultural University(Natural Science Edition)

基　　金：内蒙古自治区自然科学基金项目(2014MS0614);内蒙古自治区高等学校科学研究项目(NJZY20021)。

摘　　要：蒙古文由蒙文字元组成,针对手写体蒙文字元样本过少的问题,提出基于深度卷积对抗生成网络(DCGAN)的样本增强方法对手写体蒙文字元样本进行样本增强。DCGAN与WGAN均属于对抗生成网络中具有代表性的网络模型,实验使用DCGAN网络模型生成的图像与WGAN网络模型生成的图像进行对比,发现同轮次下DCGAN的运算速度与生成图像质量都优于WGAN。在实验中使用DCGAN对手写体蒙文字元样本进行样本增强,生成新的样本图像并对样本进行分类输入原数据集,增加样本多样化。利用颜色直方图、感知哈希算法对深度卷积对抗生成网络生成的图像进行分析评估,发现新生成的样本图像与原图像相似度高达84.8%。Mongolian is composed of Mongolian characters. Aiming at the problem of too few samples of Mongolian characters in handwriting,a sample enhancement method based on the DCGAN was proposed to enhance the samples of Mongolian characters in handwriting. Both the DCGAN and WGAN are representative network models in the generative adversarial networks. In the experiments,the images generated by the DCGAN network model were compared with those generated by the WGAN network model,and it was found that the operation speed and quality of DCGAN were better than those of WGAN in the same round. In this paper,the DCGAN network model was used to enhance the samples of Mongolian characters,generate new sample images,and input the samples into the original data set to increase the sample diversification. By using the color histogram and perceptual hash algorithm to analyze and evaluate the images generated by the network,it was found that the similarity between the new sample images and the original images was as high as 84. 8%.

关 键 词：DCGAN WGAN 样本增强 图像 分类 

分 类 号：TP391.1[自动化与计算机技术—计算机应用技术]