摘要
图像风格迁移在前馈网络中主要经过内容表征、风格表征、图像重建三步完成。针对在迁移过程中风格控制参数为零时生成图像与输入内容图像发生偏离,生成图像风格出现过拟合情况,提出一种零风格图像重建方法。通过在前馈网络的每次迭代中增加{内容图像,内容图像}数据对训练,修改AdaIN风格差值函数,重新计算相应总损失进行图像重建。实验结果显示该方法在风格迁移过程中更加稳定迅速的特点,并解决图像风格过拟合问题。
Image style transfer is mainly completed in three stages of content image representations, style image characterization and image reconstruction in feed-forward neural networks. In the process of style transfer, the generated images of the previous feed-forward networks with the style control parameter which is equal to zero, are not same to the content image but an image of some biased style. Therefore, proposes a zero- style reconstruction method to solve the over-fitting of output image style strength. In there, reconstructs output image by adding (content image, content image) data pair in each iteration of the training phase, modifies the AdaIN style difference function, and then recalculates the total style loss. The experimental results verify that the proposed method is more stable and rapid in the style transfer process, and it solves the problem of image style over-fitting.
作者
余义斌
林治
吴承鑫
YU Yi-bin;LIN Zhi;WU Chen-xin(Department of Intelligent Manufacturing, Wuyi University, Jiangnien 529020)
出处
《现代计算机》
2019年第13期49-53,58,共6页
Modern Computer
关键词
风格迁移
前馈网络
零风格图像
风格差值法
Style Transfer
Feed-Forward Networks
Zero-Style Image
Style-Interpolation Method