基于深度学习的图像风格迁移方法研究  

Research on image style transfer based on deep learning

作　　者：梁永侦 Liang Yongzhen(School of Information Engineering,Guangxi Technological College of Machinery and Electricity,Nanning,Guangxi 530007,China)

机构地区：[1]广西机电职业技术学院信息工程学院,广西南宁530007

出　　处：《计算机时代》2023年第8期107-112,共6页Computer Era

基　　金：广西机电职业技术学院科研项目(2020YKYZ003);广西高校中青年教师科研基础能力提升项目(2022KY1071)。

摘　　要：为了快速实现图像风格迁移效果,提出基于深度学习的图像风格迁移方法。先将预处理目标内容图像和风格图像输入预训练VGG-19网络进行特征提取,再对特征损失系数进行设置,进而引入权重因子和平衡损失系数的最小二乘惩罚函数以加权求和方式获取总损失系数,最后依据梯度下降迭代方法进行图像重建,实现图像风格迁移效果。实验采用峰值信噪比(PSNR)和均方误差(MSE)来评价风格迁移图像的性能。结果显示,整体取得了很好的图像风格迁移效果。In order to quickly realize the effect of image style transfer,a method based on depth learning is proposed.Firstly,the pre-processed target content image and style image are input into the pre-trained VGG-19 network for feature extraction.Then the feature loss coefficient is set,and the least squares penalty function of the weighted factor and the balance loss coefficient is introduced to obtain the total loss coefficient by weighted summation.Finally,the gradient descent iterative method is used for image reconstruction to achieve the effect of image style transfer.The PSNR and MSE are used to evaluate the performance of image style transfer.The results show that the overall image style transfer effect is very good.

关 键 词：深度学习 图像风格迁移 峰值信噪比 均方误差 

分 类 号：TP301.6[自动化与计算机技术—计算机系统结构]