欢迎访问《图学学报》 分享到:

图学学报

• 计算机视觉 • 上一篇    下一篇

基于生成模型的古壁画非规则破损部分修复方法

  

  1. (云南大学信息学院,云南 昆明 650500)
  • 出版日期:2019-10-31 发布日期:2019-11-06

The Inpainting of Irregular Damaged Areas in Ancient Murals  Using Generative Model

  1. (School of information Science and Engineering, Yunnan University, Kunming Yunnan 650500, China)
  • Online:2019-10-31 Published:2019-11-06

摘要: 为了更好地保存和修复珍贵的古代壁画艺术,在现有的人工修复技术之上,结合 数字虚拟修复方法,使用深度学习中生成网络的方法自动生成壁画缺失部分,可以有效地提高 修复效率,降低修复成本。用于修复的网络整体上是一个自编码器结构,编码器将待处理壁画 图像和破损部分对应的掩膜作为输入,进行特征提取。解码器将编码器得到的特征图通过反卷 积的方法恢复到原来尺寸,完成修复,自动将破损区域进行补全。同时,通过对待修复壁画进 行分块修复再拼接的方法实现了对任意尺寸壁画的修复。与其他数字壁画修复方法相比,该方 法更加通用,不受壁画种类和破损情况的限制。在一般破损的壁画上可以得到超过目前先进水 平的修复效果,并且在人眼无法辨识有效信息的大面积破损的壁画上,仍可以恢复得到有完整 语义的图像。

关键词: 壁画修复, 卷积神经网络, 生成模型, 自编码器

Abstract: In order to preserve and restore the precious ancient mural art in a better way, based on the existing manual restoration technology, the digital virtual restoration method can effectively improve the efficiency of restoration and reduce the costs of restoration. In this aspect, using the generative network method in deep learning to automatically generate the missing part of the murals for completion and restoration can achieve good results. The network used for restoration is basically an autoencoder. The encoder takes the murals images to be processed and the mask corresponding to the damaged part as the input for feature extraction. The decoder will restore the feature chart obtained from the encoder to its original size by deconvolution, which completes the restoration. In this process, the damaged area will be completed automatically. At the same time, separating the murals into different pieces, restoring and reassembling them later makes it achievable to restore murals of any size. Compared with other digital mural restoration methods, the one proposed in the present study is applicable to more general purposes and not limited by the type of murals and their damage. In the generally damaged murals, this method can achieve a better restoration effect compared with the existing level. Moreover, even for a large-area damaged mural where the naked eyes cannot identify effective information, this method can nevertheless restore it to one containing images of full meaning.

Key words: murals repainting, convolutional neural network, generating model, autoencoder