基于层间互相关感知损失的风格迁移方法  

Style transfer based on cross-layer correlation perceptual loss

作　　者：庄轩权 李彩霞[1] 黎培兴[1,2] ZHUANG Xuanquan;LI Caixia;LI Peixing(School of Mathematics,Sun Yat-sen University,Guangzhou 510275,China;Guangdong Province Key Laboratory of Computational Science,Sun Yat-sen University,Guangzhou 510275,China)

机构地区：[1]中山大学数学学院,广东广州510275 [2]中山大学广东省计算科学重点实验室,广东广州510275

出　　处：《中山大学学报（自然科学版）》2020年第6期126-135,共10页Acta Scientiarum Naturalium Universitatis Sunyatseni

基　　金：广东省基础与应用基础研究基金(2020B1515310007);中山大学广东省计算科学重点实验室(2020B1212060032)。

摘　　要：深度学习在风格迁移领域的应用使一系列以图片艺术风格化为核心的产品真正落地,而从像素级损失向基于Gram矩阵的感知损失的转变是其中最关键的跨越。Gram矩阵在艺术风格特征的提取上有良好的效果,但其局限于同等级语义特征间相关性统计的做法并不能作为艺术风格的充分表示。自Gram矩阵被提出以来,一系列研究并未对其进行充分的研究和改进,而是关注于模型结构的设计以提高风格迁移的速度。提出使用层间互相关矩阵作为Gram矩阵的代替或补充进行风格迁移任务的风格损失函数计算。实验表明,在得到相似水平输出结果的情况下,使用层间互相关矩阵方法可以降低20%的计算时间。Great success in deep-learning-based style transfer is accelerating the development of photo artistic stylization applications.And the change of loss function from per-pixel loss to perceptual loss based on the Gram matrix is the most critical part of this progress.Gram matrix shows good performance in style feature extraction,but it only focuses on correlations among same level features.Therefore,Gram matrix cannot be considered as a complete representation of styles.However,most of the research focus on how to improve transfer speed by designing new model structure instead of analyzing and modifying the Gram matrix.The cross-layer correlation matrix is used to calculate style loss function as a replacement or supplement to the Gram matrix.By experiments,it is shown that this method can reduce 20%of the calculation time in comparison with the Gram matrix method while yielding similar outputs.

关 键 词：风格迁移 GRAM矩阵 卷积神经网络 风格损失函数 感知损失 深度学习 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]