首页|期刊简介|投稿指南|分类索引|刊文选读|订阅指南|证明资料|样刊邮寄查询|常见问题解答|联系我们
白明明,张运杰,张膑. 利用对抗性边缘学习模型生成超分辨率图像[J]. 科学技术与工程, 2021, 21(19): 7891-7898.
Bai Mingming,Zhang Yunjie,Zhang Bin.Generating Image Super-resolution with Adversarial Edge Learning Model[J].Science Technology and Engineering,2021,21(19):7891-7898.
利用对抗性边缘学习模型生成超分辨率图像
Generating Image Super-resolution with Adversarial Edge Learning Model
投稿时间:2020-10-29  修订日期:2021-04-13
DOI:
中文关键词:  超分辨率  卷积神经网络  生成对抗网络  边缘特征学习
英文关键词:super-resolution  convolutional neural network  generative adversarial network  edge feature learning
基金项目:
        
作者单位
白明明 大连海事大学
张运杰 大连海事大学
张膑 大连海事大学
摘要点击次数: 191
全文下载次数: 71
中文摘要:
      大部分基于卷积神经网络的图像超分辨率方法都是采用端到端的模式,这类图像超分辨率方法往往存在重构图像纹理边缘模糊、高频信息缺失的问题。为了改善该问题,在SRGAN的基础上提出了一种基于对抗性图像边缘学习的深层网络模型,将图像边缘信息得到充分利用,来引导超分网络生成更加真实的高分辨率图像。该网络模型由两个生成对抗网络所组成,首先利用一个生成对抗网络来生成低分辨率图像所对应的高分辨率边缘特征图,然后再用高分辨率边缘特征图来约束和引导第二个生成对抗网络,使之重构出来的高分辨率图像纹理边缘更加清晰,更好的恢复图像边缘的高频细节。在 Set5、Set14、BSD100、Urban100和Manga109基准测试集上的实验结果表明了该算法重构出的高分辨率图像更加接近真实的图像,在峰值信噪比、结构相似度和感知指标上都有不错的表现。
英文摘要:
      Most of the image super-resolution method based on convolutional neural network adopt end-to-end mode, which often has problems in reconstructing the image texture and missing high-frequency information. In order to solve this problem, a deep network structure based on super-resolution generation adversarial networks (SRGAN) for adversarial image edge learning was proposed. The edge information of the image was used to guide the super-resolution network to generate more real high-resolution images. The network model consists of two pairs of generative adversarial networks. First of all, using a generate adversarial network to generate high resolution edge feature map of the low resolution image , and then using the edge feature map to restrain and guide the second generate adversarial network to reconstruct the high resolution image. The experimental results on Set5, Set14, BSD100, Urban100 and Mangr109 benchmark dataset demonstrate that the reconstructed high-resolution image is more close to the real image, and has a good performance in terms of peak signal-to-noise ratio, structural similarity and perception index.
查看全文  查看/发表评论  下载PDF阅读器
关闭
你是第44239227位访问者
版权所有:科学技术与工程编辑部
主管:中国科学技术协会    主办:中国技术经济学会
Tel:(010)62118920 E-mail:stae@vip.163.com
京ICP备05035734号-4
技术支持:本系统由北京勤云科技发展有限公司设计

京公网安备 11010802029091号