基于学习的人脸超分辨率重建技术研究

发布时间：2018-09-18 12:30

【摘要】：数字图像的分辨率度量了图像的清晰度,分辨率的高低是由图像单位面积上像素的个数决定的,它是衡量图像质量高低的关键指标。人脸超分辨率重建是一种对低分辨率人脸处理得到对应高分辨率人脸的低成本的软件技术,又称为人脸幻生。基于人脸图像或视频广泛涉及于计算机视觉、机器学习等领域,人脸幻生技术一直是热门关注、研究方向,它被广泛应用于多媒体通信、人脸识别、视频监控等。本文重点研究了基于学习的人脸幻生技术,通过对基于相似性约束的人脸幻生算法和基于邻域嵌入的人脸幻生算法的研究,提出改进的基于相似性约束的迭代学习人脸幻生算法和改进的基于位置的邻域嵌入人脸幻生算法。主要研究成果及贡献包括以下两方面:1.传统学习型图像超分辨率重建方法存在着高空间复杂度的缺陷,由此提出改进的基于相似性约束的迭代学习人脸幻生算法。首先让每幅待重建人脸通过主成分分析在初始训练集人脸库中筛选中与之相似度最高的若干个人脸对构成新的训练集,用于后续的迭代学习重建;每次迭代仅有一个训练集高低分辨率人脸对参与训练,降低了空间复杂度;由于流形学习中存在着低维空间到高维空间的一对多映射的缺陷,提出在每次迭代中均采用相似性约束计算重建权重,以降低这种不一致性导致的重建误差。实验结果表明,本章提出的改进算法不仅有较小的空间复杂度,同时也拥有更好的主观和客观效果。2.研究基于邻域嵌入的人脸幻生技术。同上,首先让每幅待重建人脸经过一个自适应筛选训练集相似人脸库系统的预处理,得到新的训练集人脸库。为了更好地表示人脸图像的信息,提出新的人脸特征提取方法和高低分辨率人脸块特征联合学习的邻域人脸块搜索方法;基于人脸的结构相似性,对人脸进行基于位置的人脸块分类;考虑到流形学习存在着低分辨率流形到高分辨率流形间的不一致性,对待重建位置人脸块的训练集高低分辨率邻域人脸块分别投影至共同的流形空间以进行流形校正,然后后续的重建权重系数的局部邻域嵌入算法求解均建立在共同流形上进行。实验证明,本章提出的改进的基于位置的邻域嵌入人脸幻生技术相对传统邻域嵌入算法,重建出的高分辨率人脸图像有更好的客观质量和主观效果。
[Abstract]:The resolution of the digital image measures the definition of the image, and the resolution is determined by the number of pixels in the unit area of the image, which is the key index to measure the image quality. Face super-resolution reconstruction is a kind of low-cost software technology for low-resolution face processing, which is also called face illusion generation. Based on face image or video is widely involved in computer vision, machine learning and other fields, face fantasy technology has been a hot concern, research direction, it has been widely used in multimedia communications, face recognition, video surveillance and so on. This paper focuses on the study of face illusion generation technology based on learning, through the study of face illusion generation algorithm based on similarity constraint and neighborhood embedding face fantasy generation algorithm. An improved iterative learning face illusion generation algorithm based on similarity constraints and an improved location-based neighborhood embedded face illusion generation algorithm are proposed. The main research results and contributions include the following two aspects: 1. The traditional super-resolution reconstruction method of learning image has the defect of high space complexity, so an improved iterative learning algorithm based on similarity constraints is proposed. First, each face to be reconstructed is selected by principal component analysis (PCA) in the initial training set face database to form a new training set, which can be used for the subsequent iterative learning reconstruction. There is only one training set in each iteration to participate in the training, which reduces the space complexity, because of the defect of one-to-many mapping from low-dimensional space to high-dimensional space in manifold learning. In order to reduce the reconstruction error caused by this inconsistency, similarity constraints are used to calculate the reconstruction weights in each iteration. Experimental results show that the proposed improved algorithm not only has less space complexity, but also has better subjective and objective effects. This paper studies the technology of facial illusion generation based on neighborhood embedding. First, each face to be reconstructed is preprocessed by an adaptive training set similar face database system, and a new training set face database is obtained. In order to better represent face image information, a new face feature extraction method and a neighborhood block search method based on high and low resolution face block feature learning are proposed. Face block classification based on position is carried out, considering the inconsistency between low resolution manifold and high resolution manifold in manifold learning. The training set of the training set for reconstruction position face blocks are projected to the common manifold space to correct the manifold respectively. Then the local neighborhood embedding algorithm of the subsequent reconstruction weight coefficients is established on the common manifold. Experimental results show that the improved location-based neighborhood embedding algorithm has better objective quality and subjective effect than the traditional neighborhood embedding algorithm.
【学位授予单位】：南京邮电大学
【学位级别】：硕士
【学位授予年份】：2015
【分类号】：TP391.41

【相似文献】