摘要
人脸超分辨率重建是指从一幅低分辨率人脸图像重建出相应的高分辨率图像的过程。大部分的人脸超分辨率重建算法都假设输入图像是对齐且不含噪声的。当输入的人脸图像为非对齐时,超分辨率重建的性能将降低。为此,提出一种基于学习的层级聚类回归算法,其主要针对非对齐的单帧人脸图像的超分辨率重建。该算法分为两部分:聚类和回归。聚类阶段,将训练样本的尺寸统一成某个小尺寸的人脸图像,用于训练人脸图像字典。该字典的字典原子为聚类中心,对原始的人脸图像进行聚类,得到各个子空间的人脸图像簇。该算法充分利用了人脸结构的先验信息,能获得更准确的聚类结果。在回归阶段,仅需要训练一个全局字典,各个子空间的人脸图像共享这些字典原子。在每个簇内,搜索各个驻点的邻域,以生成对应的邻域子空间。然后,学习低分辨率与高分辨率样本特征之间的映射关系,以得到每个子空间的回归模型。该算法的核心是所有的人脸图像类共享一个全局字典,但对于同一个驻点,在不同的人脸图像簇内,邻域样本各不相同,这样能够更准确地学习局部映射关系。该算法不仅可以缩短训练时间,还可以提高人脸超分辨率重建的质量。对比实验的结果表明,该算法的PSNR至少可以提升0.39 dB,SSIM可以提升0.01~0.18。
Face hallucination reconstruction refers to the process of reconstructing high-resolution enhanced face from a low-resolution image.Most of the traditional methods assume that the input image is aligned and noise-free.However,the super resolution performance will decrease when the input facial image is unaligned and affected by noise.This paper proposed an effective single image super resolution method for unaligned face images,in which the learning-based hierarchical clustering regression approach is used to get better reconstruction model.The proposed face hallucination method can be divided into clustering and regression.In the clustering part,a dictionary is trained on the whole face image with tiny size,and the training images are clustered based on the Euclidean distance.Thus,the facial structural prior is fully utilized and the accurate clustering result can be obtained.In the regression part,to reduce the time complexity effectively,only one global dictionary needs to be trained during the entire training phase whose atoms are taken as the anchors.In particular,the learned anchors are shared with all the clusters.For each cluster,the Euclidean distance is used to search the nearest neighbors for each anchor to form the subspace.Moreover,in every subspace,a regression model is learned to map the relationship between low-resolution features and high-resolution samples.The core idea of this method is to utilize the same anchors but different samples for clusters to learn the local mapping more accurately,which can reduce training time and improve reconstruction quality.The results of comparative experiments with other algorithms show that the PSNR can be increased by at least 0.39 dB and the SSIM can be increased by 0.01 to 0.18.
作者
王淑云
干宗良
刘峰
WANG Shu-yun;GAN Zong-liang;LIU Feng(Jiangsu Province Key Lab on Image Processing and Image Communication,Nanjing University of Posts and Telecommunications,Nanjing 210003,China)
出处
《计算机科学》
CSCD
北大核心
2019年第8期298-302,共5页
Computer Science
基金
国家自然科学基金(60802021,61172118,61271240,61471201)
江苏省高校自然科学重点研究项目(13KJA510004)
江苏省自然科学基金青年基金(BK20130867)
江苏省高校自然科学研究(12KJB510019)资助
关键词
超分辨率
人脸超分辨率
回归
层级
欧氏距离
Super resolution
Face hallucination
Regression
Hierarchical
Euclidean distance
