Non-local注意力机制生成对抗网络的油画修复研究  

Research on Oil Painting Restoration of Non-local Attention Mechanism Generative Adversarial Network

作　　者：何妍 何嘉[1] HE Yan;HE Jia(College of Computer Science and Technology,Chengdu University of Information Technology,Chengdu 610225,China)

机构地区：[1]成都信息工程大学计算机学院,四川成都610225

出　　处：《成都信息工程大学学报》2022年第1期34-39,共6页Journal of Chengdu University of Information Technology

基　　金：四川省科技厅苗子工程资助项目(2019Z118);四川省科技厅应用基础重点资助项目(2017JY0011)。

摘　　要：针对部分油画艺术作品存在图像破损的问题,提出一种基于非局部(Non-local)注意力机制生成对抗网络的油画修复方法。首先,在生成器部分,采用扩张卷积和门控卷积替代原网络中的普通卷积层,增强网络的特征提取能力,同时加入Non-local注意力机制,提升生成器的修复能力;其次,使用马尔科夫判别器,强化网络的判别效果;最后,在损失函数部分使用感知损失、GAN损失和L1损失,使整个网络的训练更加稳定。网络在开源的Gallerix油画数据集上进行了验证,实验结果表明:与Global&Local、ParticalConvGAN和Deepfillv2的修复方法相比,PSNR和SSIM指标均超过3种优秀的生成对抗网络,对油画艺术品的修复的效果进行了更好的提升。Aiming at the problem of image damage in some oil painting artworks,a method of oil painting restoration based on Non-local attention mechanism generative adversarial networksgeneration adversarial network is proposed.First,in the part of the generator part,dilated convolution and gated convolution are used to replace the ordinary convolution layer in the original network,to strengthen the judgment effect of the network,and a non-local attention mechanism is added to improve the repair abilityrestoration effect of the generator;secondly,Markov discrimination is used to strengthen the discriminative effect of the network;finally,the loss function uses perceptual loss,GAN loss,and L1 loss are used in the loss function to make,which makes the training of the entire network more stable.The network of this article is verified on the open source Gallerix oil painting dataset.The experimental results show that the PSNR and SSIM indicators surpass the three excellent generative adversarial networks,and improve the effect of oil painting artwork restoration.

关 键 词：生成对抗网络 NON-LOCAL 油画修复 扩张卷积 门控卷积 

分 类 号：TP751[自动化与计算机技术—检测技术与自动化装置]