基于双生成器生成对抗网络的壁画图像虚拟修复方法  

The mural image virtual restoration method based on bi-generator generative adversarial network

作　　者：杨卓林 曹建芳[2] 张英俊[1] 彭存赫 Yang Zhuolin;Cao Jianfang;Zhang Yingjun;Peng Cunhe(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,太原030024 [2]忻州师范学院计算机系,忻州034000

出　　处：《国外电子测量技术》2024年第7期191-200,共10页Foreign Electronic Measurement Technology

基　　金：国家自然科学基金面上项目(62372397);教育部人文社会科学研究项目(规划基金项目)(21YJAZH002);山西省自然基金面上项目(202203021221222);山西省文物局2024年度文物科研课题(2024KT23)项目资助。

摘　　要：针对现有的基于生成对抗网络的壁画修复方法,其生成样本缺乏多样性,容易造成大规模特征丢失等问题。提出一种基于双生成器生成对抗网络(BGGAN)的壁画图像虚拟修复方法。首先,从两个随机方向进行样本生成,保证了生成样本的多样性。其次,对Dilate U-Net Kares生成器模型,改进下采样阶段的膨胀卷积扩张率,取消池化操作。最后,设计损失函数,将均方误差(MSE)损失与对抗损失相结合,通过λG约束生成样本的特征梯度。在所收集壁画数据集上进行修复测试,测试结果与多种图像修复方法对比。结果表明,所提算法获得的图像修复结果细节更清晰。修复后图像的峰值信噪比(PSNR)相较对比模型平均提高了约1.12 dB,结构相似度(SSIM)平均提高了约0.047。Aiming at the existing fresco restoration methods based on generative adversarial networks,their generated samples lack diversity and are prone to large-scale feature loss and other problems.A virtual restoration method for fresco images based on bi-generator generative adversarial network(BGGAN)is proposed.Firstly,sample generation from two random directions ensures the diversity of generated samples.Secondly,for the Dilate U-Net Kares generator model,the inflated convolutional expansion rate in the downsampling stage is improved and the pooling operation is eliminated.Finally,the loss function is designed to combine the MSE loss with the adversarial loss,and the feature gradient of the generated samples is constrained byλG.Restoration tests are performed on the collected mural dataset,and the test results are compared with multiple image restoration methods.The results show that the image restoration results obtained by the proposed algorithm have clearer details.The peak signal-to-noise ratio(PSNR)of the restored image is improved by about 1.12 dB on average compared to the comparison model,and the structural similarity(SSIM)is improved by about 0.047 on average.

关 键 词：图像处理 壁画虚拟修复 生成对抗网络 U-Net 

分 类 号：O436[机械工程—光学工程]