摘要
利用低分辨率图像生成高分辨率图像的过程称为图像超分辨率,目的是得到一张清晰的影像。随着人工智能的蓬勃发展,在遥感、辅助文本识别等诸多领域,图像超分辨率的应用愈加广泛。本文利用生成对抗网络的深度学习模型进行单图像超分重建,SRGAN模型相较于传统方法,提出了新的感知损失函数,由对抗损失和内容损失组成。对抗损失通过训练判别器网络结构区分生成图像和实际高分辨率图像,而内容损失则利用预训练的VGG19网络模型计算图像特征的感知相似度,而不是在像素空间上的相似度。试验证明,利用SRGAN获得的高分辨率图片,MOS指标高于传统方法。本文围绕SRGAN的原理、效果、应用等进行了阐述。
The process of using low resolution(LR)images to predict the corresponding high resolution(HR)images is called image super-resolutions,which aims to get a clear image.With the vigorous development of artificial intelligence,image super-resolution reconstruction has more and more applications in many fields such as medical treatment and remote sensing.This article learns to use the deep learning model of generative adversarial networks(GAN)to perform single-image super-reconstruction.Compared with traditional methods,this model proposes a new perceptual loss function,including an adversarial loss and a content loss.The adversarial loss distinguishes the generated image from the actual high-resolution image by training the discriminator network structure,while content loss uses the pre-trained VGG19 network model to calculate the perceived similarity of image features,rather than the similarity in pixel space.The experiment proves that the introduction of generative adversarial network into single-image super-resolution,MOS score is higher than traditional methods.This article will focus on the principles,effects,and applications of SR and GAN.
作者
韩志晟
孙丕川
唐超
HAN Zhisheng;SUN Pichuan;TANG Chao(Beijing Urban Construction and Surveying Design Research Institute Co.,Ltd.,Beijing 100101,China;Beijing Key Laboratory of Deep Foundation Pit Geotechnical Engineering of Rail Transit,Beijing 100101,China)
出处
《测绘通报》
CSCD
北大核心
2021年第8期106-110,共5页
Bulletin of Surveying and Mapping
关键词
单图像超分
生成对抗网络
VGG19网络模型
内容损失函数
对抗损失函数
single image super-resolution
generative adversarial networks
VGG19 networks
content loss function
perceptual loss function
