基于注意力机制与对比学习的任意风格迁移  

Arbitrary Style Transfer Based on Attention Mechanism andContrastive Learning

作　　者：高静欣 黄海峤[2] 罗霄[1] GAO Jingxin;HUANG Haiqiao;LUO Xiao(School of Arts and Sciences,Beijing Institute of Fashion Technology,Beijing 100029,China;Research Center of Beijing Digitalization Project in Fashion Industry,Beijing Institute of Fashion Technology,Beijing 100029,China)

机构地区：[1]北京服装学院文理学院,北京100029 [2]北京服装学院北京市服装产业数字化工程技术研究中心,北京100029

出　　处：《北京服装学院学报（自然科学版）》2024年第3期78-86,95,共10页Journal of Beijing Institute of Fashion Technology：Natural Science Edition

基　　金：北京市教育委员会科学研究计划项目(KM202210012002);2023年北京高等教育“本科教学改革创新项目”(202310012004)。

摘　　要：针对目前任意风格迁移研究中存在的既要保持图像语义内容和结构信息的一致性,同时又要迁移更多风格特征的难点问题,本文基于注意力机制和对比学习设计了图像任意风格迁移模型。其中所设计的基于注意力的自适应转换网络,可以使模型自适应地获取风格特征的全局统计信息,有效改善特征融合过程中内容特征与风格特征统计特性的对齐状态;而其中所设计的对比损失,使模型在对比学习中有效融合内容图像的结构特征,使生成图像与内容图像具有良好的内容相似性。同其他5种先进模型的对比实验表明,本文设计的模型生成的图像在视觉效果以及定量评价上都具有显著优势。Addressing the challenges in arbitrary image style transfer,which require maintaining the consistency between semantic content and structural information while transferring more style features,an image arbitrary style transfer model was designed based on attention mechanism and contrastive learning.The key component of the model,the Attentional Adaptive Transformation Network,enables the model to adaptively obtain the global statistics of style features,thereby improving the alignment of content and style features during feature fusion.A contrastive loss is designed and introduced into the model training process to effectively integrate the structural features of the content image through contrastive learning,enabling the generated image to have good semantic content similarity with the content image.Experimental results show that the images generated using the proposed model have significant advantages in both visual effects and quantitative evaluation indicators compared to five other advanced models.

关 键 词：任意风格迁移 注意力机制 对比学习 卷积神经网络 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]