摘要
随着深度学习的不断发展与应用,图像到图像的风格迁移成了计算机视觉领域的研究热点之一。该文运用Cycle⁃GAN网络对图像进行风格迁移,能够在无匹配的源图像和风格图像的情况下进行。使得CycleGAN网络中的生成器由编码器、转换器及解码器组成,能起到保留原始图像特征和转换图像数据的作用。该文尝试通过Inception与ResNet进行结合,并与原CycleGAN网络进行对比训练,结果表明,修改后的CycleGAN能够比CycleGAN训练后得到更加逼真的图像,具有更佳的视觉效果。
With the development and application of deep learning,image to image style transfer has become one of the research hotspots in the field of computer vision.In this paper,we use the cyclegan network to transfer the style of image,which can be done without matching source image and style image.The generator in cyclegan network is composed of encoder,converter and decoder,which can keep the original image features and transform image data.In this paper,we try to combine perception with RESNET and train in the original cyclegan network.The results show that the modified cyclegan can get more realistic images and better visual effects than cyclegan after training.
作者
马赫
张涛
卢涵宇
MA He;ZHANG Tao;LU Han-yu(School of Surveying and Mapping Science and Engineering,Shandong University of Science and Technology,Qingdao 266000,China;Guizhou Lichuang Technology Development Co.,Ltd.,Guiyang 550018,China;College of Big Data and Information En-gineering,Guizhou University,Guiyang 550025,China)
出处
《电脑知识与技术》
2020年第27期18-20,共3页
Computer Knowledge and Technology
基金
贵州省人才平台资助项目(黔科合[2017]5305,5788)。
