文章编号：1004－5422(2021)04－0367－07

DOI:10.3969/j.issn.1004-5422.2021.04.006

基于ImageNet预训练卷积神经网络的图像风格迁移

赵卫东¹，施实伟²，周  婵³

（1.成都大学计算机学院，四川  成都  610106；2.西南大学  教育学部，重庆 400715；

3.西北师范大学  教育技术学院，甘肃  兰州  730071）

摘  要：图像风格迁移技术是对目标图像的内容进行艺术化处理，在降低艺术创作难度等领域有广泛应用前景.基于深度学习的Tensorflow框架，使用ImageNet预训练卷积神经网络提取内容特征，格拉姆矩阵计算图像风格差异，将风格损失函数和内容损失函数融合构建模型整体损失函数，并利用卷积神经网络的反向传播算法，更新优化模型参数，最终生成具既有内容图像的原始内容又具有风格图像的艺术风格的目标图像.

关键词：风格迁移；卷积神经网络；格拉姆矩阵；损失函数

中图分类号：TP183;TP391.41      文献标识码：A

Image Style Transfer Based on ImageNet Pre-Trained  Convolutional Neural Network

ZHAO Weidong¹,SHI Shiwei²,ZHOU Chan³

（1.School of Computing,Chengdu University,Chengdu 610106,China;

2.School of Education,Southwest University,Chongqing 400715,China;

3.School of Educational Technology,Northwest Normal University,Lanzhou 730070,China）

Abstract：Image style transfer technology is the artistic processing of the content of the target image,which has a wide application prospect in reducing the difficulty of artistic creation.Based on deep learning Tensorflow framework,ImageNet pretrained convolutional neural network was used to extract content features,Gram matrix was used to calculate image style differences,style loss function and content loss function were fused to construct the overall loss function of the model,and the back propagation algorithm of convolutional neural network was used to update and optimize model parameters,and generate an art style target image that has both the original content of the  content image and the style image.

Key words：style transfer; convolutional neural network,Gram matrix; loss function