摘要
近年来,稀疏表示的方法在图像超分辨率(Super-resolution,SR)重建方面取得了较好的结果.但是,由于图像在获取的过程中受外界因素的影响,获取到的低分辨率(Low Resolution,LR)图像细节往往损失严重,在图像SR重建中LR图像自身可利用的先验信息有限.因此,传统的稀疏表示不能很好地恢复出图像的高频细节.针对这一不足,本文基于稀疏表示的SR重建思想,采用邻域回归的方法从外部样例中学习图像的梯度先验信息来弥补图像自身先验不足的缺点,提出一种梯度正则化稀疏表示的图像SR重建方法.该方法通过构建一种图像梯度正则化项来引导图像的细节重建,提高重建的高分辨率(High Resolution,HR)图像的质量.实验结果表明,本文提出的SR重建算法重建结果较好,能恢复出更清晰的边缘信息,在主观和客观上重建结果都优于大多数的方法.
In recent years,the sparse representation method has achieved good results in image super-resolution(SR)reconstruction.However,due to the influence of external factors in the process of image acquisition,the low resolution(LR)image details obtained are often seriously lost,and the prior information available to LR image itself is limited in image SR reconstruction.Therefore,the traditional sparse representation cannot recover the high-frequency details of the image very well.To solve this problem,this paper uses the neighborhood regression method to learn the gradient prior information of image from external samples to make up for the shortcomings of the image’s ow n prior,and proposes an image SR reconstruction method based on the sparse representation of gradient regularization.In this method,an image gradient regularization term is constructed to guide the detail reconstruction of the image and improve the quality of HR image reconstruction.Experimental results show that the SR reconstruction algorithm proposed in this paper has better reconstruction results and can recover more clear edge information.M eanw hile,both the subjective and objective reconstruction results of the proposed method are superior to most SR methods.
作者
黄淑英
胡晓燕
吴昕
吴佳俊
许亚婷
HUANG Shu-ying;HU Xiao-yan;WU Xin;WU Jia-jun;XU Ya-ting(School of Software&Internet of Things Engineering,Jiangxi University of Finance and Economics,Nanchang 330032,China)
出处
《小型微型计算机系统》
CSCD
北大核心
2020年第12期2588-2594,共7页
Journal of Chinese Computer Systems
基金
国家自然科学基金项目(61862030,61662026,62072218)资助
江西省自然科(20182BCB22006,20181BAB202010,20192ACB20002,20192ACBL21008)资助。
关键词
图像超分辨率
稀疏表示
梯度正则化
邻域回归
image super-resolution
sparse representation
gradient regularization
neighborhood regression
