基于VGG-19和MMD卷积神经网络模型的国画风格迁移
Style migration of Chinese painting based on VGG-19 and MMD convolutional neural network models
徐子俊 1胡予昕 1陆文浩 1宋兴睿 1刘哲1
作者信息
- 1. 江汉大学人工智能学院,武汉 430056
- 折叠
摘要
卷积神经网络因效果强大而被广泛应用于图像识别,在提取图像特征方面有极大的进步.由于风格迁移技术主要是针对西方油画,而国画是一种传统的中国艺术风格,其在风格迁移方向上缺乏广泛的应用.设计以国画代替西方油画作为风格图像,以自然景观照片作为内容图像,探究传统国画经过卷积神经网络后的提取效果.实验依据VGG算法模型并结合TensorFlow 2框架,对采集的数据集进行预处理,采集像素制成数据矩阵,输入VGG-19浅层模型进行训练,通过MMD最小化分布特征图差异,增强卷积层的目标效果.该方法取得比较满意的结果,可为风格迁移转换的研究提供更多参考.
Abstract
The convolutional neural network(CNN)has been widely used in image recognition due to its powerful capabilities in extracting image features.However,style transfer techniques have primarily focused on Western oil paintings,lacking extensive applications in the context of traditional Chinese art styles,such as Chinese ink paintings.In this study,we aimed to explore the ef-fect of convolutional neural networks on extracting features from traditional Chinese ink paintings by replacing Western oil paint-ings with Chinese ink paintings as style images and using natural landscape photographs as content images.We conducted experi-ments based on the VGG algorithm model and TensorFlow 2 framework,preprocessing the collected dataset by converting pixel val-ues into data matrices.These matrices were then fed into the VGG-19 shallow model for training.We further minimized the differ-ences in distribution feature maps using the Maximum Mean Discrepancy(MMD)technique to enhance the target effect of the con-volutional layers.This approach yielded satisfactory results and can provide valuable insights for further research on style transfer transformations.
关键词
卷积层神经网络/VGG-19/MMD/风格迁移算法Key words
convolution layer neural network/VGG-19/MMD/style transfer algorithm引用本文复制引用
基金项目
湖北省大学生创新创业训练计划(S202211072096)
武汉市科技局知识创新专项曙光计划(2022010801020382)
武汉市教育局市属高校教研课题(2021016)
江汉大学校级科研项目四新专项(2022SXZX17)
江汉大学高层次人才科研启动经费项目(2019025)
出版年
2024
NETL
NSTL
万方数据