摘要
图像到图像转换是一类视觉和图形问题,目标是通过使用一组配对的图像来学习输入图像和输出图像之间的映射.然而,对于许多任务来说,配对的训练数据难以获得.CycleGAN提出了一种在没有成对示例的情况下学习从源域x到目标域y的图像转换的方法.一般来说,生成式对抗网络(GAN)在训练的过程中,由生成器产生的图片将不再能够欺骗判别器,判别器容易战胜生成器,因此平衡生成器和判别器的训练稳定程度有时比性能指标更重要.为了解决此问题,本文将训练数据的分类标签加入GAN的训练,提出了一种改进的CycleGAN半监督模型,以解决GAN训练不稳定问题.实验结果表明,生成的图像显示更加真实,具有更佳视觉效果.
Image-to-image translation is a class of vision and graphics problems where the goal is to learn the mapping between an input image and an output image using a training set of aligned image pairs.However,for many tasks,paired training data will not be available.CycleGAN present an approach for learning to translate an image from a source domain X to a target domain Y in the absence of paired examples.Basically,the GAN discriminator easily overwhelms the generator.If it happens,the generator tries to fool the discriminator in an improper way.Thebalance between discriminator and generator is important for the performance.To solve this problem,smooth labeling is used and an improved CycleGAN semi-supervised model architecture is changed and solved the problem of GAN training instability.Our proposed method performs better visual improvements in our results which are morerealistic.
作者
何剑华
龙法宁
朱晓姝
HE Jian-hua;LONG Fa-ning;ZHU Xiao-sbu(Educational Technology Center,Yulin Normal University,Yulin,Guangxi 537000;College of Computer Science,Yulin Normal University,Yulin,Guangxi 537000)
出处
《玉林师范学院学报》
2018年第2期122-126,共5页
Journal of Yulin Normal University
基金
广西重点实验室科研课题项目(2016CSOBDP0302)
广西高校科研项目(2013YB202)
