基于改进CycleGAN的人脸卡通风格化迁移
Face cartoon style transfer based on improved CycleGAN
杜润梅 1李旭辉 1刘铭1
作者信息
- 1. 长春工业大学 数学与统计学院,吉林 长春 130012
- 折叠
摘要
利用改进的循环生成对抗网络 CycleGAN 实现人脸卡通风格化迁移.文中模型在CycleGAN基础上通过对生成器模型进行结构改进,在编码器部分采用稠密卷积结构,使模型在减少了参数量的同时可以更好地关注人脸细节特征,加强特征传播,在不改变个人脸型的基础上实现人脸卡通风格化迁移.实验结果表明,改进后的模型风格化迁移图像分辨率更高,配色更协调,尤其是细节处如眼睛、发丝等卡通迁移效果更流畅.
Abstract
Face cartoon stylised transfer using improved CycleGAN.The model in this paper is based on CycleGAN by making structural improvements to the generator model.A dense convolutional structure is used in the encoder part,so that the model can better focus on the detailed features of the face and enhance the feature propagation while reducing the number of parameters.Achieve cartoon stylised transfer of faces without changing the individual's face shape.The results of the comparison experiments show that the improved model stylized transfer image has higher resolution and more coordinated colour scheme,especially the details such as eyes,hair and other cartoon transfer effect is smoother.
关键词
风格迁移/CycleGAN/DenseNet模型/编码器Key words
style transfer/CycleGAN/DenseNet model/encoder引用本文复制引用
基金项目
吉林省发改委基本建设资金项目(2022C043-2)
吉林省自然科学基金项目(20200201157JC)
出版年
2024
NETL
NSTL
万方数据