一种基于有限数据的改进DCGAN图像生成方法

摘要：

生成对抗网络(Generative Adversarial Network,GAN)的成功主要依赖于大量的数据进行模型训练.当训练数据有限时,GAN生成图像会产生保真度低和模型不稳定等问题.针对以上问题,基于深度卷积生成对抗网络(Deep Convolutional Generative Adversarial Networks,DCGAN)提出一种改进模型,称为LC-DCGAN(LeCam Deep Convolutional Generative Adversarial Networks),通过引入两个指数移动平均变量,减少小批量之间的方差,并且来稳定正则化项,使其判别器的预测逐渐收敛到平稳点.实验结果表明,该模型在有限数据下可以生成高质量、缺陷类型丰富的缺陷样本数据集.

The success of Generative Adversarial Network(GAN)mainly relies on a large amount of data for model training.When the training data is limited,GAN has problems such as unstable model and low fidelity of the generated image.To solve the above problems,an improved model based on Deep Convolutional Generative Adversarial Networks(DCGAN)is proposed in this paper,called LC-DCGAN(LeCam Deep Convolutional Generative Adversarial Networks).By introducing two exponential moving average variables to reduce the variance between small batches and stabilize the regularization term,so that its discriminator prediction gradually converges to the stable point.The experimental results indicate that the model can generate high-quality and diverse defect sample datasets under limited data.

作者：

王士斌 高梓雕 刘栋

Wang Shibin;Gao Zidiao;Liu Dong(.Computer and Information Engineering College,Key Laboratory of Henan Province,Henan Normal University,Xinxiang 453007,China;“Educational Artificial Intelligence and Personalized Learning”,Key Laboratory of Henan Province,Henan Normal University,Xinxiang 453007,China)

机构地区：

betway官方app 计算机与信息工程学院

出处：

《betway官方app 学报:自然科学版》 CAS 北大核心 2023年第6期39-46,共8页

Journal of Henan Normal University(Natural Science Edition)

基金：

国家自然科学基金(62072160) 河南省科技攻关计划项目(222102210187).

关键词：

图像生成 生成对抗网络 正则化 保真度

image generation Generative Adversarial Networks regularization fidelity

分类号：

TP399 [自动化与计算机技术—计算机应用技术]