低算力深度学习下的图像卡通风格化研究  

Cartoon style transfer of images with low performance deep learning

作　　者：徐鹏飞[1] 周腾骅 武仲科[1] 申佳丽[1] 王醒策[1] XU Pengfei;ZHOU Tenghua;WU Zhongke;SHEN Jiali;WANG Xingce(School of Artificial Intelligence,Beijing Normal University,Beijing Key Laboratory of Digital Preservation and Virtual Reality for Cultural Heritage,Engineering Research Center of Virtual Reality and Applications,Ministry of Education:100875,Beijing,China)

机构地区：[1]北京师范大学人工智能学院,文化遗产数字化保护与虚拟现实北京市重点实验室,虚拟现实应用教育部工程研究中心,北京100875

出　　处：《北京师范大学学报（自然科学版）》2021年第6期888-895,共8页Journal of Beijing Normal University(Natural Science)

基　　金：国家重点研发计划资助项目(2020YFC1523302,2017YFB1002604,2017YFB1402105);国家重点研发金砖国家合作资助项目(2017YFE0100500);国家自然科学基金资助项目(61972041,62072045)。

摘　　要：在深入研究图像风格迁移的基础上,提出了一种适用于图形处理器性能受限情况下,卡通(cartoon)图像风格迁移训练的生成式对抗网络(generative adversarial networks,GAN).利用视觉几何组(visual geometry group,VGG)网络提取图片先验信息,实现学习过程的加速;裁剪cartoonGAN模型,在保证效果的基础上,使得低性能计算条件下的网络收敛成为可能;设计合理的损失函数,保证整体风格化效果.基于tensorflow 2.0构建试验平台,通过对试验结果分析可发现,该方法的迁移效果好,稳定性强,且收敛时间短.对算法的参数和初始化方法给出了相关讨论,并提出了进一步的解决方案.The technology of image style transfer has become increasingly important in computer vision,having achieved amazing results.However,since cartoon style is different from most art styles,there is still ample room for further improvements.This paper reviews relevant background,highlights significance of image style transfer,then focuses on cartoonGAN with limited GPU performance.Image priori information was extracted from VGG network to accelerate the learning process.CartoonGAN model was tailored to make convergence possible under low performance computing conditions due to guaranteed results.Reasonable loss function was designed to ensure overall styling effect.The network was implemented under open source deep learning framework tensorflow 2.0.Data analysis and possible improvement methods are presented.

关 键 词：风格迁移 深度学习 生成式对抗网络 卡通风格化 

分 类 号：TP231[自动化与计算机技术—检测技术与自动化装置]