摘要
针对现有算法在修复大面积破损区域的壁画图像时容易出现特征提取困难,上下文结构不一致等问题,提出一种基于双判别生成对抗网络的壁画图像虚拟修复。首先,将U-Net架构引入生成网络,结合扩张卷积与跳跃连接实现多尺度特征融合提取,利用重构损失初步构建修复模型。其次利用双重判别网络,保证图像全局一致性的同时,加强修复后的局部细节。最后交替训练生成网络和双重判别网络,加权重构损失和WGAN-GP损失,进一步优化网络模型,完成破损壁画图像的虚拟修复。根据创建的壁画数据集,进行训练测试,并与多组修复算法进行修复对比,结合主客观评价指标进行评价,结果表明,该算法修复的壁画图像质量更优,较好的完成了较大区域受损壁画图像的整体一致性修复。
Aiming at the problems of difficulty in feature extraction and inconsistent context structure when existing algorithms repair mural images in large damaged areas,a virtual restoration of mural images based on bi-discriminative generative adversarial network is proposed.First of all,the U-Net architecture is introduced into the generation network,combined with dilated convolution and skip connections to achieve multi-scale feature fusion extraction,and the reconstruction loss is used to initially build a repair model,and then the double discriminant network is used to ensure the global consistency of the image and at the same time strengthen the local details after repairing.Finally,the generation network and the double discriminant network are alternately trained,and the weighted reconstruction loss and WGAN-GP loss are used to further optimize the network model and complete the virtual restoration of damaged mural images.According to the created mural data set,carry out training and testing,and compare the restoration with multiple groups of restoration algorithms,and evaluate based on subjective and objective evaluation indicators.The results show that the mural image quality restored by this algorithm is better,and the larger area is better completed.Overall consistency fixes for damaged mural images.
作者
胡雅妮
李光亚
韩晓东
简丽
张国花
Hu Yani;Li Guangya;Han Xiaodong;Jian Li;Zhang Guohua(School of Information and Communication Engineering,North University of China,Taiyuan 030051,China;Shanxi Research Institute for the Protection of Ancient Architecture and Painted Mural Paintings,Taiyuan 030000,China)
出处
《国外电子测量技术》
北大核心
2022年第6期14-19,共6页
Foreign Electronic Measurement Technology
基金
国家重点研发计划(2020YFB2009102)项目资助。
关键词
壁画虚拟修复
生成对抗网络
U-Net
扩张卷积
virtual restoration of murals
generative adversarial network
U-Net
dilated convolution
