摘要
生成对抗网络已经成为深度学习领域最热门的研究方向之一,其最大的优势在于能够以无监督的方式来拟合一个未知的分布。目前,生成对抗网络在图像生成领域大放异彩,其能够产生一些高质量的图像,但也暴露了一些弊端。在生成图像的过程中,经常会出现模式坍塌问题,从而导致生成的样本过于单一。为了解决这个问题,对生成对抗网络的模型结构和损失函数加以改进,使判别器能够从多个角度来度量生成数据的分布和真实数据的分布之间的差异,从而改善生成样本的多样性。通过在多个数据集上进行实验,结果显示,提出的模型在很大程度上缓解了模式坍塌问题。
Generative adversarial networks have become one of the most popular research directions in the field of deep lear-ning.Its main advantage is that it can fit unknown distribution in an unsupervised way.At present,the generative adversarial network is valuable in the field of image generation.It can generate some high-quality images,but it also exposes some disadvantages.In the process of image generation,the problem of mode collapse often occurs,which leads to the generated sample being too single.To solve this problem,this paper improved the model structure and loss function of the generative adversarial network,so that the discriminator could measure the difference of distribution between generated data and real data from many aspects,thus increasing the diversity of generated samples.Experiments on multiple data sets show that the proposed model alleviates the mode collapse to a large extent.
作者
尹来国
孙仁诚
邵峰晶
隋毅
邢彤彤
Yin Laiguo;Sun Rencheng;Shao Fengjing;Sui Yi;Xing Tongtong(Dept.of Computer Science&Technology,Qingdao University,Qingdao Shandong 266071,China)
出处
《计算机应用研究》
CSCD
北大核心
2022年第6期1689-1693,共5页
Application Research of Computers
基金
国家自然科学青年基金资助项目(41706198)。
关键词
生成对抗网络
图像生成
模式坍塌
generative adversarial network
image generation
mode collapse
