南京邮电大学学报（自然科学版）2025，Vol.45Issue(2)：93-102,10.DOI:10.14132/j.cnki.1673-5439.2025.02.011
基于快速自适应元学习的小样本学习
Few-shot learning based on fast adaptive meta-learning
摘要
Abstract
Common generative adversarial network(GAN)synthesize new,realistic images through ad-versarial learning,but they require a large amount of training data.Inspired by the human brains'ability to quickly learn new concepts from a few examples,this paper proposes a fast-adaptive meta-learning model based on GANs and encoder networks for few-shot image generation.This model only requires a small number of examples and can generate images of unseen target categories by training a simplified network and increasing the number of generator iterations.Compared to the comparative models,it con-verges five times faster and reduces the number of trainable parameters needed to a quarter.Experimental results show that the fast-adaptive meta-learning model has the highest image quality,diversity,and clar-ity in few-shot image generation,and its outputs can reach a level of fidelity comparable to images in com-mon datasets.The proposed method can effectively enhance the performance of few-shot image generation.关键词
小样本学习/元学习/图像生成技术/无监督学习/生成对抗网络Key words
few-shot learning/meta-learning/image generation techniques/unsupervised learning/gen-erative adversarial network(GAN)分类
信息技术与安全科学引用本文复制引用
马涛,赵华,樊卫东,罗华峰,吴强,石瑞达,张铁勋..基于快速自适应元学习的小样本学习[J].南京邮电大学学报（自然科学版）,2025,45(2):93-102,10.基金项目
国家电网有限公司总部科技项目(5700-202319302A-1-1-ZN)资助项目 （5700-202319302A-1-1-ZN）
