基于改进生成对抗网络的图像风格迁移方法研究
Research on image style transfer method based on improved generative adversarial network
司周永 1王军号1
作者信息
- 1. 安徽理工大学 计算机科学与工程学院,安徽 淮南 232001
- 折叠
摘要
为了解决传统GAN(Generative Adversarial Network)进行图像风格迁移受到成对数据集的限制,以及Cycle-GAN学习高级特征时表现不佳和训练过慢的问题,本文采用ModileNetV2-CycleGAN模型进行图像风格迁移,并引入多尺度结构相似性指数(multi-scale structural similarity,MS-SSIM)作为惩罚项保留风格图片的特征,来提高特征学习的效果,从而提高风格化图片质量.采用客观结构相似性SSIM与峰值信噪比PSNR和主观投票作为评估指标,对迁移后的效果进行评估,实验结果表明了本文改进算法的有效性.
Abstract
In order to solve the problem of the traditional Generative Adversarial Network image-style transfer limited to the pairing data set,and the problems of poor performance and slow training when CycleGAN learn advanced features,the paper uses the ModileNetv2-CycleGAN model for image style transfer,and introduces multiscale structural similarity index loss as a characteristic of punishment items retains style pictures to improve the effect of characteristic learning,thereby improving the quality of style pictures.Objective structure similar to SSIM and peak signal-to-noise ratio PSNR and subjective voting as evalu-ation indicators are adopted to evaluate the effect of transfer.The experimental results show that using ModileNetv2-CycleGAN and MS-SSIM Loss can improve style migration quality and have better visual effects.
关键词
图像风格迁移/循环一致性生成对抗网络/轻量级卷积神经网络/深度残差网络/多尺度结构相似性指数Key words
image style transfer/cycle consistent generative adversarial networks/lightweight convolution neural network/deep residual network/multi-scale structural similarity引用本文复制引用
出版年
2024
NETL
NSTL
万方数据