基于卷积神经网络的风格迁移艺术字研究

Research on Wordart of Style Transfer Based on Convolutional Neural Network

针对艺术字风格迁移只迁移风格图像的颜色特征、生成字形风格单一的问题,提出了一种基于卷积神经网络(Convolutional Neural Network,CNN)的风格迁移艺术字的方法。该方法首先通过字库提取多种类型字体,自动生成内容图像,再经过预训练VGG19网络提取风格图像的抽象特征表示。构造Gram矩阵作为图像风格表征,最后利用L-BFGS算法进行迭代优化,生成具有特殊风格的艺术字体。结果与市面上艺术字生成器产生的艺术字进行对比,本文的风格迁移艺术字兼具其纹理特征和颜色特征,更具有美感。

Aiming at the problems that the style transfer of wordart only transfers the color features of style images and generates a single font style, this paper proposes a method of style transfer of wordart based on convolutional neural networks. This method first extracts multiple types of fonts from the font library, automatically generates content images, extracts the abstract feature representation of style images through the pre-trained VGG19 network, and constructs a Gram matrix as the image style representation. Finally, L-BFGS is used for iterative optimization to generate artistic fonts with special styles. Compared with the characters made by wordart generators on the market, the style transfer wordarts in this paper are more aesthetic, with both texture and color features.