基于深度学习的风格迁移方法综述  

A Review of Style Transfer Methods Based on Deep Learning

作　　者：刘嘉雄 周骏[1] LIU Jiaxiong;ZHOU Jun(School of Computer and Information Science/School of Software,Southwest University,Chongqing,400715,China)

机构地区：[1]西南大学计算机与信息科学学院/软件学院,重庆400715

出　　处：《人工智能科学与工程》2024年第1期1-19,共19页Journal of Southwest China Normal University(Natural Science Edition)

基　　金：国家自然科学基金面上项目(22274134)。

摘　　要：随着深度学习的迅猛发展,风格迁移技术在算法和应用上取得了重大突破,为内容与风格的创新交互提供了强大支持。该文综述了风格迁移的基本概念、分类及其在神经网络中的应用,特别是神经网络风格迁移的原理、变体与合成算法。文章还对基于文本的图像风格迁移与基于图像的方法进行了比较,分析了各自的优缺点,突显了智能化风格迁移技术的发展。此外,探讨了风格迁移技术与其他领域结合的情况,如与超分辨方法和对比学习方法等的结合,以及在大型工艺品设计中的应用实例,展示了其广泛的应用潜力。该文的目的是为研究者提供清晰的视角,推动风格迁移领域的技术进步。With the rapid development of deep learning,style transfer technology has made significant breakthroughs in algorithms and applications,providing strong support for innovative interaction between content and styles.This paper reviews the basic concepts,classifications and applications of style transfer in neural networks,focusing on the principles,variations and synthesis algorithms of neural style transfer.A comparison between text-based image style transfer and image-based methods was conducted,analyzed their respective advantages and disadvantages,highlighted the development of intelligent style transfer technology.Furthermore,the integration of style transfer technology with other fields was discussed,such as its combinations with super-resolution methods and contrastive learning methods,as well as its application examples in large-scale artwork design,demonstrated its extensive potential applications.The purpose of this paper was providing researchers with a clear perspective and promoting technological advancement in the field of style transfer.

关 键 词：神经风格迁移 深度学习 算法 性能提升 基于文本的风格迁移 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]