摘要
针对传统图像修复算法存在的图像细节修复效果差、视觉连贯性不佳以及训练不稳定等问题,将生成对抗网络和孪生神经网络进行结合,孪生神经网络当成GAN中的判别器,并在生成网络中使用均方误差,孪生网络中使用对比损失.实验结果显示,SN-GAN模型在细节修复及视觉连贯性上得到一定提升,并且也更适用于大面积缺损图像的修复.
Aiming at the problems of poor image detail restoration,poor visual coherence and unstable training of traditional image restoration algorithms,the generative adversarial network and siamese network were combined,and the siamese network was treated as a discriminator in GAN.The mean square error was used in the generative network and the contrast loss in the siamese network.The experimental results show that the SN-GAN model gets some improvement in detail restoration and visual coherence,and is also more applicable to the restoration of large defective images.
作者
贺佳馨
吕晓琪
张继凯
李菁
HE Jiaxin;LYU Xiaoqi;ZHANG Jikai;LI Jing(Information Engineering School,Inner Mongolia University of Science and Technology,Baotou 014010,China;Institute of Information Engineering,Inner Mongolia University of Technology,Hohhot 010051,China)
出处
《内蒙古科技大学学报》
CAS
2022年第2期180-186,共7页
Journal of Inner Mongolia University of Science and Technology
基金
国家自然科学基金资助项目(61771266)
内蒙古自治区自然科学基金资助项目(2019BS06005)
内蒙古自治区高等学校科学研究基金资助项目(NJZY20095)
内蒙古自治区科技计划基金资助项目(2019GG138)。
关键词
图像修复
生成对抗网络
孪生神经网络
缺损图像
image inpainting
generative adversarial networks
siamese network
defect image
