Face Hallucination: Theory and Practice

In this paper, we study face hallucination, or synthesizing a high-resolution face image from an input low-resolution image, with the help of a large collection of other high-resolution face images. Our theoretical contribution is a two-step statistical modeling approach that integrates both a global parametric model and a local nonparametric model. At the first step, we derive a global linear model to learn the relationship between the high-resolution face images and their smoothed and down-sampled lower resolution ones. At the second step, we model the residue between an original high-resolution image and the reconstructed high-resolution image after applying the learned linear model by a patch-based non-parametric Markov network to capture the high-frequency content. By integrating both global and local models, we can generate photorealistic face images. A practical contribution is a robust warping algorithm to align the low-resolution face images to obtain good hallucination results. The effectiveness of our approach is demonstrated by extensive experiments generating high-quality hallucinated face images from low-resolution input with no manual alignment.     

结合位置先验与稀疏表示的单帧人脸图像超分辨率算法

提出了一种结合位置先验与稀疏表示的人脸图像超分辨率算法,可对单帧输入的低分辨率人脸图像基于训练集进行超分辨率重建。利用压缩感知理论中的信号分解方法,〖BP(〗明确哪些方法更好〖BP)〗,将稀疏表示与人脸位置先验信息相结合,使用经过分类的超完备冗余字典,来分别稀疏逼近输入信号的块向量结构。利用最佳的K项原子,线性组合重建出高分辨率图像块。最后按照图像块最初在人脸的位置,将它们拼接为整体人脸。在CAS-PEAL-R1人脸图库上的实验结果表明,该算法使用相对较少的原子,就可以重建出质量较好的高分辨率人脸图像。     

Parts-based face super-resolution via non-negative matrix factorization

Face super-resolution refers to inferring the high-resolution face image from its low-resolution one. In this paper, we propose a parts-based face hallucination framework which consists of global face reconstruction and residue compensation. In the first phase, correlation-constrained non-negative matrix factorization (CCNMF) algorithm combines non-negative matrix factorization and canonical correlation analysis to hallucinate the global high-resolution face. In the second phase, the High-dimensional Coupled NMF (HCNMF) algorithm is used to compensate the error residue in hallucinated images. The proposed CCNMF algorithm can generate global face more similar to the ground truth face by learning a parts-based local representation of facial images; while the HCNMF can learn the relation between high-resolution residue and low-resolution residue to better preserve high frequency details. The experimental results validate the effectiveness of our method.     

A two-step face hallucination approach for video surveillance applications

Multimedia Tools and Applications - In this paper we propose a novel face hallucination algorithm to synthesize a high-resolution face image from several low-resolution input face images. Face...     

基于低分辨率彩色指导图像的深度图像超分辨率重建

传统的以彩色图像为指导的深度图像超分辨率(SR)重建方法,参考图像必须为高分 辨率彩色图像,彩色图像的分辨率决定了深度图像的放大上限。同时,实际应用中可能只存在低 分辨率彩色图像,此时上述方法也不再适用。为此,探讨使用任意分辨率彩色图像为指导的深度 图像SR 重建方法。首先,使用大量不同类别的图像SR 算法对输入彩色图像进行上采样,得到 高分辨率彩色图像并以此作为指导图像,然后采用基于二阶总广义变分方法,将由低分辨率彩色 图像重建得到的图像作为正则约束项,添加图像边缘信息,构建目标函数,将深度图像SR 重建 问题转化为最优化问题,再通过原-对偶方法求解,最终得到高分辨率深度图像。探讨了之前被 相关方法所忽略的情形,该方法可以适用于任意分辨率的彩色指导图像。并且通过相关实验发现 了令人惊异的现象,即通过使用低分辨率彩色图像放大后作为指导,可以得到与使用高分辨率彩 色指导图像相近甚至更好的结果,对相关问题的研究和应用具有一定参考意义。     

多尺度密集残差网络的单幅图像超分辨率重建

目的 近几年应用在单幅图像超分辨率重建上的深度学习算法都是使用单种尺度的卷积核提取低分辨率图像的特征信息,这样很容易造成细节信息的遗漏。另外,为了获得更好的图像超分辨率重建效果,网络模型也不断被加深,伴随而来的梯度消失问题会使得训练时间延长,难度加大。针对当前存在的超分辨率重建中的问题,本文结合GoogleNet思想、残差网络思想和密集型卷积网络思想,提出一种多尺度密集残差网络模型。方法 本文使用3种不同尺度卷积核对输入的低分辨率图像进行卷积处理,采集不同卷积核下的底层特征,这样可以较多地提取低分辨率图像中的细节信息,有利于图像恢复。再将采集的特征信息输入残差块中,每个残差块都包含了多个由卷积层和激活层构成的特征提取单元。另外,每个特征提取单元的输出都会通过短路径连接到下一个特征提取单元。短路径连接可以有效地缓解梯度消失现象,加强特征传播,促进特征再利用。接下来,融合3种卷积核提取的特征信息,经过降维处理后与3×3像素的卷积核提取的特征信息相加形成全局残差学习。最后经过重建层,得到清晰的高分辨率图像。整个训练过程中,一幅输入的低分辨率图像对应着一幅高分辨率图像标签,这种端到端的学习方法使得训练更加迅速。结果 本文使用两个客观评价标准PSNR（peak signal-to-noise ratio）和SSIM（structural similarity index）对实验的效果图进行测试,并与其他主流的方法进行对比。最终的结果显示,本文算法在Set5等多个测试数据集中的表现相比于插值法和SRCNN算法,在放大3倍时效果提升约3.4 dB和1.1 dB,在放大4倍时提升约3.5 dB和1.4 dB。结论 实验数据以及效果图证明本文算法能够较好地恢复低分辨率图像的边缘和纹理信息。     

Global consistency,local sparsity and pixel correlation: A unified framework for face hallucination

In this paper, a novel two-phase framework is presented to deal with the face hallucination problem. In the first phase, an initial high-resolution (HR) face image is produced in patch-wise. Each input low-resolution (LR) patch is represented as a linear combination of training patches and the corresponding HR patch is estimated by the same combination coefficients. Realizing that training patches similar with the input may provide more appropriate textures in the reconstruction, we regularize the combination coefficients by a weighted _?2-norm${?}_2\mathrm{-}\mathrm{norm}$ minimization term which enlarges the coefficients for relevant patches. The HR face image is then initialized by integrating all the HR patches. In the second phase, three regularization models are introduced to produce the final HR face image. Different from most previous approaches which consider global and local priors separately, the proposed algorithm incorporates the global reconstruction model, the local sparsity model and the pixel correlation model into a unified regularization framework. Initializing the regularization problem with the HR image obtained in the first phase, the final output HR image can be optimized through an iterative procedure. Experimental results show that the proposed algorithm achieves better performances in both reconstruction error and visual quality.     

基于特征空间的人脸超分辨率重构

超分辨率图像重构是利用关于同一场景的多帧低分辨率图像重构出一幅具有更高分辨率图像的过程.传统的超分辨率图像重构算法是基于像素空间,通过利用高、低分辨率像素空间之间的映射关系来求解,具有计算复杂性高等缺点. 针对低分辨率人脸放大问题,提出了一个基于特征空间的人脸超分辨率图像重构算法.与传统算法相比,该算法不仅降低了计算复杂性,还具有更好的鲁棒性.     

Locally affine patch mapping and global refinement for image super-resolution

This paper deals with the super-resolution (SR) problem based on a single low-resolution (LR) image. Inspired by the local tangent space alignment algorithm in [16] for nonlinear dimensionality reduction of manifolds, we propose a novel patch-learning method using locally affine patch mapping (LAPM) to solve the SR problem. This approach maps the patch manifold of low-resolution image to the patch manifold of the corresponding high-resolution (HR) image. This patch mapping is learned by a training set of pairs of LR/HR images, utilizing the affine equivalence between the local low-dimensional coordinates of the two manifolds. The latent HR image of the input (an LR image) is estimated by the HR patches which are generated by the proposed patch mapping on the LR patches of the input. We also give a simple analysis of the reconstruction errors of the algorithm LAPM. Furthermore we propose a global refinement technique to improve the estimated HR image. Numerical results are given to show the efficiency of our proposed methods by comparing these methods with other existing algorithms.     

基于自适应权值滤波的深度图像超分辨率重建

目的 越来越多的应用都依赖于对真实场景深度图像的准确且快速的观测和分析。飞行时间相机可以实时获取场景的深度图像,但是由于硬件条件的限制,采集的深度图像分辨率较低,无法满足实际应用的需要。为此提出一种结合同场景彩色图像通过构造自适应权值滤波器对深度图像进行超分辨率重建的方法。方法 充分发掘深度图像的非局部以及局部自相似性先验约束,结合同场景的高分辨率彩色图像构造非局部及局部的自适应权值滤波算法对深度图像进行超分辨率重建。具体来说,首先利用非局部滤波算法来有效避免重建结果的振铃效应,然后利用局部滤波算法进一步提升重建的深度图像质量。结果 实验结果表明,无论在客观指标还是视觉效果上,基于自适应权值滤波的超分辨率重建算法较其他算法都可以得到更好的结果,尤其当初始的低分辨率深度图像质量较差的情况下,本文方法的优势更加明显,峰值信噪比可以得到1dB的提升。结论 结合非局部和局部自相似性先验约束,结合同场景的高分辨率彩色图像构造的自适应权值滤波算法,较其他算法可以得到更理想的结果。     

Multi-resolution feature fusion for face recognition

For face recognition, image features are first extracted and then matched to those features in a gallery set. The amount of information and the effectiveness of the features used will determine the recognition performance. In this paper, we propose a novel face recognition approach using information about face images at higher and lower resolutions so as to enhance the information content of the features that are extracted and combined at different resolutions. As the features from different resolutions should closely correlate with each other, we employ the cascaded generalized canonical correlation analysis (GCCA) to fuse the information to form a single feature vector for face recognition. To improve the performance and efficiency, we also employ “Gabor-feature hallucination”, which predicts the high-resolution (HR) Gabor features from the Gabor features of a face image directly by local linear regression. We also extend the algorithm to low-resolution (LR) face recognition, in which the medium-resolution (MR) and HR Gabor features of a LR input image are estimated directly. The LR Gabor features and the predicted MR and HR Gabor features are then fused using GCCA for LR face recognition. Our algorithm can avoid having to perform the interpolation/super-resolution of face images and having to extract HR Gabor features. Experimental results show that the proposed methods have a superior recognition rate and are more efficient than traditional methods.     

A novel adaptive image zooming scheme via weighted least-squares estimation

A critical issue in image interpolation is preserving edge detail and texture information in images when zooming. In this paper, we propose a novel adaptive image zooming algorithm using weighted least-square estimation that can achieve arbitrary integer-ratio zoom (WLS-AIZ) For a given zooming ratio n, every pixel in a low-resolution (LR) image is associated with an n × n block of high-resolution (HR) pixels in the HR image. In WLS-AIZ, the LR image is interpolated using the bilinear method in advance. Model parameters of every n × n block are worked out throughweighted least-square estimation. Subsequently, each pixel in the n × n block is substituted by a combination of its eight neighboring HR pixels using estimated parameters. Finally, a refinement strategy is adopted to obtain the ultimate HR pixel values. The proposed algorithm has significant adaptability to local image structure. Extensive experiments comparingWLS-AIZ with other state of the art image zooming methods demonstrate the superiority of WLS-AIZ. In terms of peak signal to noise ratio (PSNR), structural similarity index (SSIM) and feature similarity index (FSIM), WLS-AIZ produces better results than all other image integer-ratio zoom algorithms.     

Nuclear norm regularized coding with local position-patch and nonlocal similarity for face hallucination

Learning-based face hallucination methods have received much attention and progress in past few decades. Specially, position-patch based approaches have been proposed to replace the probabilistic graph-based or manifold learning-based ones. As opposed to the existing patch based methods, where the input image patch matrix is converted into vectors before combination coefficients calculation, in this paper, we propose to directly use the image matrix based regression model for combination coefficients computation to preserve the essential structural information of the input patch matrix. For each input low-resolution (LR) patch matrix, its combination coefficients over the training image patch matrices at the same position can be computed. Then the corresponding high-resolution (HR) patch matrix can be obtained with the LR training patches replaced by the corresponding HR ones. The nonlocal self-similarities are finally utilized to further improve the hallucination performance. Various experimental results on standard face databases indicate that our proposed method outperforms some state-of-the-art algorithms in terms of both visual quantity and objective metrics.     

低分辨率图像的妆容迁移算法

现有的妆容迁移算法效果优越, 功能丰富, 但是较少考虑到输入图像为低分辨率的场景. 当高分辨率图像难以获得时, 现有的妆容迁移算法将难以适用, 妆容无法完全迁移. 为此本文提出了一种适用于低分辨率图像的妆容迁移算法, 将包含妆容信息的特征矩阵作为先验信息, 将超分辨率网络与妆容迁移网络结合在一起产生协同效应, 即使输入的图像为低分辨率图像, 也能输出高分辨率的妆容迁移结果, 并且充分保留妆容细节的同时提升姿势和表情的鲁棒性. 由于使用端到端的模型实现妆容迁移与超分辨率, 因此设计了一组联合损失函数, 包括生成对抗损失、感知损失、循环一致性损失、妆容损失和均方误差损失函数. 所提出的模型在妆容迁移与超分辨率的定性实验和定量实验中均达到了先进水平.     

A novel kernel-based framework for facial-image hallucination

In this paper, we present a kernel-based eigentransformation framework to hallucinate the high-resolution (HR) facial image of a low-resolution (LR) input. The eigentransformation method is a linear subspace approach, which represents an image as a linear combination of training samples. Consequently, those novel facial appearances not included in the training samples cannot be super-resolved properly. To solve this problem, we devise a kernel-based extension of the eigentransformation method, which takes higher-order statistics of the image data into account. To generate HR face images with higher fidelity, the HR face image reconstructed using this kernel-based eigentransformation method is treated as an initial estimation of the target HR face. The corresponding high-frequency components of this estimation are extracted to form a prior in the maximum a posteriori (MAP) formulation of the SR problem so as to derive the final reconstruction result. We have evaluated our proposed method using different kernels and configurations, and have compared these performances with some current SR algorithms. Experimental results show that our kernel-based framework, along with a proper kernel, can produce good HR facial images in terms of both visual quality and reconstruction errors.     

基于边缘学习的低分辨率图像识别算法

由于受光照条件、拍摄角度、传输设备以及周围环境的影响,刑侦视频图像中的目标物体往往分辨率较低,难以识别。针对低分辨率图像识别问题,在经典LeNet-5识别网络的基础上,提出了一种基于边缘学习的低分辨率图像识别算法。首先由边缘生成对抗网络生成低分辨率图像的幻想边缘,该边缘与高分辨率图像边缘相近;再将该低分辨图像的生成边缘信息作为先验信息融合到识别网络中对低分辨率图像进行识别。在MNIST、EMNIST和Fashion-mnist三个数据集上分别进行实验,结果表明,将低分辨图像的幻想边缘信息融合到识别网络中可以提高低分辨率图像的识别率。     

Morphable model space based face super-resolution reconstruction and recognition

Super-resolution image reconstruction is the process of producing a high-resolution image from a set of low-resolution images of the same scene. For the applications of performing face evaluation and/or recognition from low-resolution video surveillance, in the past, super-resolution image reconstruction was mainly used as a separate preprocessing step to obtain a high-resolution image in the pixel domain that is later passed to a face feature extraction and recognition algorithm. Such three-stage approach suffers a high degree of computational complexity. A low-dimensional morphable model space based face super-resolution reconstruction and recognition algorithm is proposed in this paper. The approach tries to construct the high-resolution information both required by reconstruction and recognition directly in the low dimensional feature space. We show that comparing with generic pixel domain algorithms, the proposed approach is more robust and more computationally efficient.     

A novel correspondence-based face-hallucination method

This paper addresses the problem of estimating high-resolution (HR) facial images from a single low-resolution (LR) input. We assume that the input LR and estimated HR images are under the same view-point and illumination condition, i.e. the setting of image super-resolution. At the core of our techniques is that the facial images can be decomposed as a texture vector, characterized in terms of the appearance, and a shape vector, characterized in terms of the geometry variations. This enables a two-stage successive estimation framework that is geometry aware and obviates the needs in sophisticated optimizations. In particular, the proposed technique first solves for appearance of the HR faces form the correspondence derived between an interpolated LR face and its corresponding HR face. Given the texture of the HR faces, we incorporate optical flow to solve the local structure at sub-pixel level for the HR faces; here, we use additional geometry inspired priors to further regularize the solution. Experimental results show that our method outperforms other state-of-the-art methods in terms of retaining the facial-feature shape and the estimation of novel features.     

Wavelet-based image interpolation using multilayer perceptrons

Changing the resolution of digital images and video is needed image processing systems. In this paper, we present nonlinear interpolation schemes for still image resolution enhancement. The proposed neural network interpolation method is based on wavelet reconstruction. With the wavelet decomposition, the image signals can be divided into several time–frequency portions. In this work, the wavelet decomposition signal is used to train the neural networks. The pixels in the low-resolution image are used as the input signal of the neural network to estimate all the wavelet sub-images of the corresponding high-resolution image. The image of increased resolution is finally produced by the synthesis procedure of wavelet transform. In the simulation, the proposed method obtains much better performance than other traditional methods. Moreover, the easy implementation and high flexibility of the proposed algorithm also make it applicable to various other related problems.     

Hallucinating faces: LPH super-resolution and neighbor reconstruction for residue compensation

A two-phase face hallucination approach is proposed in this paper to infer high-resolution face image from the low-resolution observation based on a set of training image pairs. The proposed locality preserving hallucination (LPH) algorithm combines locality preserving projection (LPP) and radial basis function (RBF) regression together to hallucinate the global high-resolution face. Furthermore, in order to compensate the inferred global face with detailed inartificial facial features, the neighbor reconstruction based face residue hallucination is used. Compared with existing approaches, the proposed LPH algorithm can generate global face more similar to the ground truth face efficiently, moreover, the patch structure and search strategy carefully designed for the neighbor reconstruction algorithm greatly reduce the computational complexity without diminishing the quality of high-resolution face detail. The details of synthetic high-resolution face are further improved by a global linear smoother. Experiments indicate that our approach can synthesize distinct high-resolution faces with various facial appearances such as facial expressions, eyeglasses efficiently.