一种渐进式增长条件生成对抗网络模型  

A progressive growing of conditional generative adversarial networks model

作　　者：马辉 王瑞琴[1] 杨帅 MA Hui;WANG Ruiqin;YANG Shuai(Huzhou University,Huzhou 313000,China)

机构地区：[1]湖州师范学院,浙江湖州313000

出　　处：《电信科学》2023年第6期105-113,共9页Telecommunications Science

基　　金：国家自然科学基金资助项目(No.62277016)。

摘　　要：渐进式增长生成对抗网络(PGGAN)是一种能够生成高分辨图像的网络模型,但是当样本间的类别不平衡或者样本类别过于相似或不相似时,容易出现模式崩溃现象而导致生成效果不佳。提出一种渐进式增长条件生成对抗网络(PGCGAN)模型,将条件生成对抗网络的思想引入PGGAN,在PGGAN的基础上加入类别信息作为条件,在网络结构和小批量标准差两个方面对PGGAN进行了改进,缓解图像生成过程中的模式崩溃现象。在对3个数据集的实验中,相比于PGGAN,PGCGAN在起始分数(IS)和Fréchet距离(FID)两个评价图像生成的指标方面都有较大程度的提升,生成的图像具有更高的多样性和真实性;且PGCGAN可以同时训练多个无关联的数据集而不崩溃,在类别不平衡或数据过于相似和不相似的数据集中均能产生高质量的图像。Progressive growing of generative adversarial networks(PGGAN)is an adversarial network model that can generate high-resolution images.However,when the categories of samples are unbalanced,or the categories of samples are too similar or too dissimilar,it is prone to produce mode collapse,resulting in poor image generation effect.A progressive growing of conditional generative adversarial networks(PGCGAN)model was proposed.The idea of conditional generative adversarial networks(CGAN)was introduced into PGGAN.Using category information as condition,PGGAN was improved in two aspects of network structure and mini-batch standard deviation,and the phenomenon of model collapse in the process of image generation was alleviated.In the experiments on the three data sets,compared with PGGAN,PGCGAN has a greater degree of improvement in inception score and Fréchet inception distance,two evaluation indicators for image generation,and the generated images have higher diversity and authenticity;and PGCGAN multiple unrelated datasets can be trained simultaneously without crashing,and high-quality images can be produced in datasets with imbalanced categories or data that are too similar and dissimilar.

关 键 词：生成对抗网络 渐进式增长条件生成对抗网络 小批量标准差 图像生成 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]