基于生成对抗网络的小样本图像数据增强技术  

Few-shot image data augmentation based on generative adversarial networks

作　　者：杨鹏坤 李金龙[1] 郝润来 Yang Pengkun;Li Jinlong;Hao Runlai(School of Computer Science and Technology,University of Science and Technology of China,Hefei 230026,China)

机构地区：[1]中国科学技术大学计算机科学与技术学院,安徽合肥230026

出　　处：《网络安全与数据治理》2023年第6期79-84,102,共7页CYBER SECURITY AND DATA GOVERNANCE

摘　　要：基于生成对抗网络(GANs)的图像数据增强方法在近年来展现出了巨大的潜力。然而生成高分辨率、高保真图像通常需要大量训练数据,这和缺乏训练数据的现状背道而驰。为解决这一问题,提出了一种能够在小样本、高分辨率图像数据集上稳定训练的条件生成对抗网络模型,并且将该模型用于数据增强。实验结果表明,在基准数据集上,该模型与当前最新模型相比能够生成更加逼真的图像并取得了最低的FID值;在图像分类任务中使用其进行数据增强能够有效缓解分类器的过拟合问题。In recent years,image data augmentation methods based on Generative Adversarial Networks(GANs)have shown great potential.However,generating high-resolution,high-fidelity images typically requires a large amount of training data,which contradicts the current lack of training data situation.To address this issue,a conditional GAN model that can stably train on fewshot,high-resolution image datasets has been proposed for data augmentation.Experimental results on benchmark datasets indicate that this model,compared to the current state-of-the-art models,is capable of generating more realistic images and achieving the lowest Fréchet Inception Distance(FID)score.Furthermore,using this model for data augmentation in image classification tasks effectively mitigates overfitting issues in classifiers.

关 键 词：生成对抗网络 数据增广 图像分类 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]