利用特征变换实现神经风格迁移的绘画机器人

李思佳¹，袁羽齐²，陈雯柏³，张玉洁¹

1. 北京邮电大学信息与通信工程学院，北京市，102206；

2. 北京邮电大学国际学院，北京市，102206；

3. 北京信息科技大学自动化学院，北京市，100192

摘要：本文主要提出了一种基于VGG网络模型的通用图像风格迁移方法，该方法的主要贡献在于，它不需要对任意一种风格进行提前学习和预训练。该算法主要通过训练重构自动解码器的方法来实现风格迁移任务。在特征提取的过程中，该算法通过白化和着色两种方式对内容特征和风格特征进行匹配。在实验中，我们对一系列摄影图像进行了测试。从实验结果可以看出，我们的方法训练过程简单、计算复杂度低、计算速度较高且对计算设备的算力要求不高，风格迁移效果良好，且具有优良的泛化能力。

关键词：卷积神经网络；风格迁移；深度学习；VGG模型；Gram矩阵；白化；着色；特征变换；风格图像

Painting robot for transforming neural style usingfeature transformation

Li Sijia¹，Yuan Yuqi²，Chen Wenbai³，Zhang Yujie¹

1. Beijing University of Post and Telecommunication, School of information and communication engineering, Beijing City, 102206;

2. Beijing University of Post and Telecommunication, International School, Beijing City, 102206;

3. Beijing Information Science and Technology University, School of Automation, Beijing City, 100192

Abstract：This paper mainly proposes a general image style migration method based on VGG network model. The main contribution of this method is that it does not need to learn and pre-train any kind of style in advance. The algorithm mainly implements the style migration task by training the method of reconstructing the automatic decoder. In the process of feature extraction, the algorithm matches the content features and style features through whitening and coloring. In the experiment, we tested a series of photographic images. The experimental results show that our method has a simple training process, low computational complexity, high computational speed and low computational power for computing equipment. Style migration works well and has excellent generalization ability.

Keywords：Convolutional neural network;Style migration;Deep learning;VGG model;Gram matrix;Whitening;Coloring;Feature transformation;Style image