SN-CLPGAN:基于谱归一化的中国传统山水画风格迁移方法  

SN-CLPGAN: A method of style transfer in Chinese traditional landscape painting based on spectral normalization

作　　者：胡琦瑶 刘乾龙 彭先霖 张翔[1] 彭盛霖 范建平[1,5] HU Qiyao;LIU Qianlong;PENG Xianlin;ZHANG Xiang;PENG Shenglin;FAN Jianping(School of Information Science and Technology,Northwest University,Xi’an 710127,China;Shaanxi Key Laboratory of Higher Education Institution of Generative Artificial Intelligence and Mixed Reality,Xi’an 710127,China;Network and Data Center,Northwest University,Xi’an 710127,China;School of Art,Northwest University,Xi’an 710127,China;State-Province Joint Engineering and Research Center of Advanced Networking and Intelligent Information Services,Xi’an 710127,China)

机构地区：[1]西北大学信息科学与技术学院,陕西西安710127 [2]生成式人工智能与混合现实陕西省高等学校重点实验室,陕西西安710127 [3]西北大学网络与数据中心,陕西西安710127 [4]西北大学艺术学院,陕西西安710127 [5]西北大学新型网络智能信息服务国家地方联合工程研究中心,陕西西安710127

出　　处：《西北大学学报(自然科学版)》2025年第1期63-74,共12页Journal of Northwest University(Natural Science Edition)

基　　金：国家自然科学基金(62471390、62306237);陕西省重点研发计划(2024GX-YBXM-149);西北大学研究生创新项目(CX2024204、CX2024206)。

摘　　要：中国传统山水画的风格迁移为文化遗产数字化保护与传承提供了新的路径,近年来,深度学习技术已实现了不同图像间的风格迁移,并展现出栩栩如生的效果。中国传统山水画的风格迁移旨在继承中国古代画家独特的绘画技巧,但存在3个缺陷:①缺乏高质量的中国传统山水画图像数据集;②忽略了中国传统山水画特有的技法和笔墨细节;③风格迁移效果与真实山水画有所差距。为了弥补上述缺陷,首先,创建了一个基于风格迁移的中国传统山水画数据集STCLP,包含4281幅高质量的中国山水画以及自然景观图像,并提出了一种基于谱归一化的中国山水画风格迁移方法SN-CLPGAN。其次,提出了在生成器中使用融合反射填充层的残差稠密块(residual-in-residual dense block,RRDB)学习中国山水画独特的笔触和技法。接着,引入了多尺度结构相似性指数测量(multi-scale structural similarity index measure,MS-SSIM)损失以减少2幅图像之间的像素差异,使生成图像更接近传统绘画的色彩和颜料。最后,采用了融合谱归一化(spectral normalization,SN)的U-Net判别器增强图像纹理细节,并确保了模型训练过程的稳定性。大量实验验证了提出的方法在中国传统山水画风格迁移任务中的有效性和先进性。The style tranfer of Chinese landscape paintings offering new avenues for the digital preservation and inheritance of cultural heritage.In recent years,deep learning technologies have enabled style transfer between different images,achieving lifelike effects.Style transfer in Chinese landscape paintings aims to preserve the unique paintings skills of ancient Chinese painters,but faces three main challenges:①The lack of high-quality datasets of traditional Chinese landscape paintings.②The oversight of the unique techniques and ink details specific to traditional Chinese landscape paintings.③The gap between the style transfer outcomes and real landscape paintings.To address these deficiencies,this paper first introduces a Chinese landscape paintings dataset for style transfer,STCLP,which contains 4281 high-quality images of Chinese landscape paintings and natural landscapes.A generative adversarial network of style transfer in Chinese landscape painting based on spectral normalization is proposed,termed SN-CLPGAN.Additionally,it introduces the use of residual-in-residual dense blocks(RRDB)with reflect padding layers in the generator to learn the distinctive brushstrokes and techniques of Chinese landscape paintings.Furthermore,it employs the multi-scale structural similarity index measure(MS-SSIM)loss to minimize pixel-level differences between images,thereby producing images closer to traditional paintings in terms of color and pigmentation.Finally,the U-Net discriminator fused with SN is utilized to enhance the textural details of images,ensuring the stability of the model training process.Extensive experiments validate the effectiveness and advancement of the proposed method in the task of style transfer for Chinese landscape paintings.

关 键 词：风格迁移 人工智能艺术 中国传统山水画 生成对抗网络 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]