基于多尺度结构自相似性的超分辨率算法  

Super Resolution Algorithm Based on Multi-scale Structural Self-similarity

作　　者：卢紫微[1,3] 吴成东 陈东岳[1] 于晓升 LU Zi-wei;WU Cheng-dong;CHEN Dong-yue;YU Xiao-sheng(College of Information Science and Engineering,Northeastern University,ShenYang 110819,China;Faculty of Robot Science and Engineering,Northeastern University,ShenYang 110819,China;School of Computer and Communication Engineering,Liaoning Shihua University,Fushun 113001,China)

机构地区：[1]东北大学信息科学与工程学院,辽宁沈阳110819 [2]东北大学机器人科学与工程学院,辽宁沈阳110819 [3]辽宁石油化工大学计算机与通信工程学院,辽宁抚顺113001

出　　处：《控制工程》2020年第5期776-780,共5页Control Engineering of China

基　　金：国家自然科学基金项目(61701101,61603080,U1713216,61702247)。

摘　　要：多尺度结构自相似性是指同一幅图像中存在相同尺度或不同尺度的相似结构,这种图像结构自相似性广泛存在于自然图像中。提出了一种基于多尺度结构自相似性的单幅图像超分辨率(Super Resolution,SR)算法,该算法不依赖于外界图像,仅在原始图像的多尺度图像中搜索低分辨率(Low Resolution,LR)图像块的最相似子块,并结合脊回归算法获得低分辨率图像块和相应高分辨率(High Resolution,HR)图像块的映射关系。此外,将原始图像进行旋转、翻转等操作,扩大内部图像块的样本空间。大量的对比实验表明,本文所提算法有效地提高了峰值信噪比(Peak Signal to Noise Ratio,PSNR)和图像可视效果。Multi-scale structural self-similarity refers to those similar structures recurring many times inside the same image,both within the same scale,as well as across different scales.In this paper,a single image super resolution(SR)method based on multi-scale structure self-similarity is proposed,which utilizes the input image itself without depending on extrinsic set of training images.Instead,the most similar image patches are searched to each low resolution(LR)image patch,from the multi-scale image of the input image,and combine with ridge regression algorithm to acquire the mapping relationship between input LR image patch and the corresponding high resolution(HR)image patch.Moreover,the rotated and flipped version of the input image is exploited to extend the internal patch sample space.Experimental results show that this method effectively improves peak signal to noise ratio(PSNR)and acquires better visual effects in reconstructed image.

关 键 词：超分辨率 结构自相似性 多尺度 脊回归 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]