基于生成对抗网络的风格化书法图像生成

Generation of Stylized Calligraphic Image Based on Generative Adversarial Network

中文关键词: 书法图像  生成对抗网络  形状特征  风格特征  颜色空间转换

英文关键词: calligraphic character  GAN  shape information  stylistic feature  color space conversion

中文摘要:

      目的 为修复书法图像中的残缺字体，提出一种提取书法图像特征并自动生成风格化书法图像的方法。方法 首先针对书法作品的灰度图像利用变分自编码器提取字体的形状特征，同时将书法图像转换至Lab颜色空间中，通过四层卷积神经网络模型对L通道进行深度学习，提取书法字体中的风格特征；然后将风格特征作为条件输入，与形状特征一起输入生成对抗网络的生成器中进行联合训练，使生成的字体带有特定风格。研究过程中构建一个包含不同风格书法字体的中国书法字体生成数据集（CCGD-2019）用作模型训练。结果 提出了一种基于变分自编码与生成对抗网络的书法字体图像生成模型，能够从标准字体或随机噪声自动生成风格化书法字体图像。结论 人眼主观评价及Fréchet初始距离计算结果表明，生成字体的识别率和视觉质量均达到了令人满意的效果。

英文摘要:

      The work aims to propose a method of extracting calligraphic image features and automatically generating stylized calligraphic image. Firstly, the variational auto-encoder was used to extract shape information of the character for the grayscale image of calligraphic works. At the same time, the calligraphic image was converted into the Lab color space and the stylistic feature was extracted through 4-layers convolutional neural network in L channel. Then, the stylistic feature was input to generator as conditions. Finally, the stylistic feature and shape information were used for joint training in generative adversarial network, which could generate Chinese characters with specific style. In addition, Calligraphy Character Generation Dataset (CCGD-2019) which contained several types of calligraphy character was constructed for model training in the process of experiment. A calligraphic character image generation model based on variational auto-encoding and generative adversarial network was proposed, which could automatically generate stylized calligraphic character images from standard characters or random noise. The results of subjective evaluation and Fréchet Inception Distance indicate that the recognition rate and the visual effect of generated characters reach the satisfactory effects.