基于循环生成对抗网络的壁画色彩修复算法  

Mural color restoration algorithm based on cyclic generative adversarial network

作　　者：曹建芳 靳梦燕[1] 李朝霞 陈泽宇[1] 马尚 CAO Jianfang;JIN Mengyan;LI Zhaoxia;CHEN Zeyu;MA Shang(College of Computer Science and Technology,Taiyuan University of Science and Technology,Taiyuan 030024,China;Department of Computer Science and Technology,Xinzhou Normal University,Xinzhou 034000,China)

机构地区：[1]太原科技大学计算机科学与技术学院,山西太原030024 [2]忻州师范学院计算机系,山西忻州034000

出　　处：《山东科技大学学报（自然科学版）》2023年第4期101-112,共12页Journal of Shandong University of Science and Technology(Natural Science)

基　　金：教育部人文社会科学研究项目(21YJAZH002);山西省高等学校人文社会科学重点研究基地项目(20190130)。

摘　　要：针对敦煌唐代壁画修复所面临的褪、变色以及修复后的壁画图像色彩存在假色和伪影的问题,提出基于循环生成对抗网络和多尺度融合协调注意力机制的壁画色彩修复算法。首先在循环一致性损失中添加同一映射损失,然后改进协调注意力机制,提出多尺度融合的协调注意力机制,最后在生成器中引入多尺度融合的协调注意力机制,对图像进行卷积核大小为1×1、3×3、5×5、7×7的多尺度卷积运算,提高生成图像的协调性。实验结果表明,与CycleGAN、WGAN等经典算法相比,本文算法在构造的壁画数据集上精度更高,可以在不依赖专家知识的情况下修复褪色壁画图像的颜色。Aiming at the problems of fading,discoloration and false color and artifacts in the restored mural images in Dunhuang murals in the Tang Dynasty,a mural color restoration algorithm based on cyclic generative adversarial network and multi-scale fusion coordinated attention mechanism was proposed.First,the same mapping loss was added to the cycle consistency loss.Then,the coordinated attention mechanism was improved and a coordinated attention mechanism of multi-scale fusion was proposed.Finally,the coordinated attention mechanism of multi-scale fusion was introduced into the generator toimprove the coordination of the generative images by multi-scale convolution operations with kernel sizes of 1×1,3×3,5×5,and 7×7.The experimental results show that the proposed algorithm achieves better accuracy than the classical algorithms such as CycleGAN and WGAN on the constructed mural datasetand can restore the color of faded mural images without relying on expert knowledge.

关 键 词：循环生成对抗网络 风格迁移 壁画色彩修复 同一映射损失 协调注意力机制 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]