基于深度学习的图像超分辨率研究
2021-11-28陈锦伦王勇王瑛
陈锦伦 王勇 王瑛


摘要:图像超分辨率是由低分辨率图像重建生成高分辨率图像的过程,是计算机视觉领域的一个研究热点。近年来,随着图像超分辨率技术理论的不断创新,从传统的插值法、重构法发展到主流的深度学习算法。文中从图像超分辨率的定义出发,梳理了图像超分辨率各个时期的代表性算法,详细介绍了基于卷积神经网络、残差网络与生成对抗网络三个主流的超分辨率模型,并讨论了各个模型的网络结构、学习策略以及损失函数等问题。最后,对图像超分辨率当前的研究情况进行总结。
关键词: 图像超分辨率; 深度学习; 神经网络; 计算机视觉
中图分类号:TP18 文献标识码:A
文章编号:1009-3044(2021)30-0024-02
开放科学(资源服务)标识码(OSID):
Image Super Resolution Based on Deep Learning
CHEN Jin-lun,WANG Yong,WANG Ying
(School of Computer, Guangdong University of Technology, Guangzhou 510006, China)
Abstract:Image super resolution is the process of generating high resolution image from low resolution image. It is a research hotspot in the field of computer vision. In recent years, with the continuous innovation of the technical theory of image super resolution, the traditional interpolation and reconstruction methods have developed to the mainstream algorithms based on deep learning. In this paper, starting from the definition of image super-resolution, representative algorithms of each period are sorted out.Then, three main super-resolution models based on convolutional neural network, residual network and generative adversarial network are introduced in detail.Also, the network structure, learning strategy and loss function of each model are discussed in particular. Finally, the development of image super-resolution is summarized.
Key words:image super-resolution; deep learning; neural network; computer vision
1 引言
近年來,随着智能手机、平板电脑的普及,人们对高分辨率图像的需求日益扩大,如何获取高质量图像越来越受到关注。在图像形成过程中,环境噪声、欠采样、光学模糊、运动模糊等因素都会造成图像成像质量较差、分辨率较低[1]。在实际应用中,由于受到开发成本、工艺水平等原因,单纯通过改善硬件设备来提高图像分辨率是不现实的。为了解决这一问题,图像超分辨率(Super-resolution)[2]技术应运而生,通过计算机软件算法提高分辨率,获取高质量图像。
图像超分辨率是将给定的一张低分辨率(LR)图像恢复为对应的具有更高视觉质量的高分辨率(HR)图像的任务。高分辨率图像细节清晰、色彩丰富,被广泛应用于安全监控、医疗成像、卫星图像、视频直播等[3]领域。……
