一种基于生成对抗网络的轻量级图像翻译模型  

A Lightweight Image-to-image Translation Model Based on GAN

作　　者：王荣达 刘宁钟[1] 李强懿 沈家全 WANG Rong-da;LIU Ning-zhong;LI Qiang-yi;SHEN Jia-quan(School of Computer Science and Technology,Nanjing University of Aeronautics and Astronautics,Nanjing 211106,China)

机构地区：[1]南京航空航天大学计算机科学与技术学院,江苏南京211106

出　　处：《计算机技术与发展》2021年第11期52-57,共6页Computer Technology and Development

基　　金：国家自然科学基金(61375021)

摘　　要：图像到图像翻译是一类视觉和图形问题,其目标是使用一组图像来学习输入图像和输出图像之间的映射。近年来关于图像到图像翻译任务的研究工作有很多,例如pix2pix和CycleGAN都在这个任务上表现得非常优秀。但是很少有研究工作是关于如何对这类图像到图像翻译模型进行压缩,降低参数量的使用。目前轻量级网络结构的设计更多是用于基于卷积神经网络的目标检测模型和分类模型,而用于生成对抗网络模型的轻量级网络结构的研究还比较少。所以,文中提出了一种基于生成对抗网络的轻量级网络结构,用于对自然图像进行图像到图像翻译,例如将图像中的马转换为斑马,或者将图像从夏天的景色转换为秋天的景色。实验结果表明,通过使用轻量级的网络结构,该方法在速度和准确性等性能指标上获得了良好的表现。Image-to-image translation is a type of visual and graphic problem whose goal is to use a set of images for learning the mapping between input and output images.In recent years,there have been a lot of research works on image-to-image translation tasks.For example,pix2pix and CycleGAN have both shown excellent results on this task.But few research works are focus on how to compress image-to-image translation models or reduce the parameters of the model.At present,the design of lightweight network structures is usually used for object detection tasks and classification task,and the research on lightweight network structures for generative adversarial nets is still relatively rare.So we propose a lightweight network structure based on generative adversarial nets for image-to-image translation tasks,such as converting horses in images to zebras or converting the scene of the image from summer to autumn.The experiment shows that the proposed method can achieve excellent performance in both speed and accuracy by using a lightweight network structure.

关 键 词：图像到图像翻译 卷积神经网络 模型压缩 轻量级网络 生成对抗网络 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]