基于风格迁移的碑帖书法虚拟修复研究

项目以文化遗产数字化保护为应用背景，以古代碑帖书法图像为研究对象，针对碑帖由于自然破坏和人为损坏所带来的斑点、裂痕及文字残缺问题，致力于研究并提出碑帖书法图像去噪方法、残缺碑帖自动生成方法以及碑帖背景风格迁移方法，探索迁移学习、深度学习在碑帖书法虚拟修复中的应用。项目通过引入卷积神经网络，并结合剪切引导滤波，在保证笔锋细节信息不丢失的前提下去除碑帖书法图像噪声；通过引入生成对抗网络，构造碑帖内容特征和风格特征分离的网络模型，并结合迁移学习训练模型，通过字体风格迁移实现残缺书法字自动生成；通过引入循环生成对抗网络，进行背景风格迁移，实现书法字体到碑帖的背景风格转换。项目将通过西安碑林博物馆碑帖信息验证提出的方法的有效性。其科学实质是在碑帖虚拟修复中引入迁移学习、深度学习，探索神经网络模型与书法之间的内在规律，为文化遗产数字化保护及书法的发扬和传承提供理论支撑和应用参考。

This project is on the background of application of digital cultural heritage protection using ancient inscription rubbing calligraphy images as the research object. To deal with spots, cracks and incomplete calligraphy characters of the inscription rubbing due to the destruction of natural and artificial damage, this research aims to study and put forward the inscription rubbing image denoising method, the incomplete calligraphy characters automatic generation method and the inscription rubbing background style transfer method, and explores the application of transfer learning and deep learning in virtual restoration of inscription rubbing. Through the introduction of convolutional neural network and the combination of shearing and guided filtering, the project can eliminate the noise of the inscription rubbing calligraphy image on the premise of not losing the detailed information of the brush edge. By introducing the generative adversarial networks, a network model of separating the content features and style features of tablets will be constructed and trained with transfer learning, and the automatic generation of missing calligraphy characters will be realized through calligraphy character style transfer. By introducing the cycle consistent generative adversarial networks, the calligraphy characters will be transferred to the background style of the calligraphy. and the virtual restoration of the calligraphy will be realized. We are planning to verify the practicability and efficiency of the solutions with the works of inscription rubbing in Forest of Steles Museum. The scientific essence of this project is to introduce transfer learning and deep learnng in virtual restoration of inscription rubbing and explore the intrinsic relationship between neural network model and calligraphy characters. The research findings of this project will provide theoretical support and application reference for the digital protection of cultural heritage and the promotion and inheritance of calligraphy.