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

作　　者：司周永 王军号[1] SI Zhouyong;WANG Junhao(School of Computer Science and Engineering,Anhui University of Science and Technology,Huainan Anhui 232001,China)

机构地区：[1]安徽理工大学计算机科学与工程学院,安徽淮南232001

出　　处：《阜阳师范大学学报（自然科学版）》2024年第2期30-37,共8页Journal of Fuyang Normal University:Natural Science

基　　金：国家自然科学基金项目(61300001)资助。

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CycleGAN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。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 evaluation 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 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]