自外而内的单幅图像超分辨率复原算法  

Single Image Super-Resolution Restoration Algorithm from External Example to Internal Self-Similarity

作　　者：郑向涛[1,2] 袁媛[1] 卢孝强[1] 

机构地区：[1]中国科学院西安光学精密机械研究所瞬态光学与光子技术国家重点实验室光学影像分析与学习中心,陕西西安710119 [2]中国科学院大学,北京100049

出　　处：《光学学报》2017年第3期57-63,共7页Acta Optica Sinica

基　　金：国家973计划(2013CB336500;2012CB719905);国家自然科学基金(61232010;61472413);中国科学院重点部署项目(KGZD-EW-T03);中国科学院光谱成像技术重点实验室开放基金(LSIT201408)

摘　　要：单幅图像超分辨率(SR)复原是一个病态逆问题,需要利用图像的先验知识进行正则化约束。提出了一种同时考虑外在样例和内在自相似性的单幅图像SR复原算法,其中外在先验知识是通过卷积神经网络从外在低分辨率-高分辨率图像对学习得到的,而内在先验约束由聚类和低秩近似实现。实验结果表明,本方法在复原效果和稳健性方面优于已有方法。Single image super-resolution（SR）restoration is an ill-posed inverse problem,in which regularization restriction is done with image priori knowledge.One single image SR method is proposed which simultaneously taking external example and internal self-similarity into account.Here the external knowledge is learned by convolutional neural network from external low-resolution-high-resolution image pairs,while the internal prior is utilized by cluster and low-rank approximation. The experimental results show that the proposed method outperforms many other existing super-resolution methods in recovery effect and robustness.

关 键 词：图像处理 超分辨率 样例学习 深度卷积网络 自相似 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]