CN102902966A

CN102902966A - Super-resolution face recognition method based on deep belief networks

Info

Publication number: CN102902966A
Application number: CN2012103875044A
Authority: CN
Inventors: 樊鑫; 林妙真
Original assignee: Dalian University of Technology
Current assignee: Dalian University of Technology
Priority date: 2012-10-12
Filing date: 2012-10-12
Publication date: 2013-01-30

Abstract

A super-resolution face recognition method based on a deep trust network relates to the field of face recognition technology. From the perspective of cognition, the present invention considers that there is an intrinsic relationship between the high and low resolution human face images corresponding to each other. However, previous studies have shown that the use of linear approximation to express this intrinsic correlation effect is restricted by linear approximation. Therefore, this internal correlation is considered to be non-linear. In view of the excellent performance of the artificial neural network on the nonlinear classification problem, the present invention adopts the neural network algorithm to capture the nonlinear association of the high and low resolution face images corresponding to each other under the attitude change. Theoretical studies and neurophysiological studies have shown that to build an intelligent processing system, it is necessary to build a deep structure, such as a system built by multi-layer nonlinear processing units. The present invention utilizes deep belief networks to mine the shared non-linear structure existing in the corresponding high and low resolution face images.

Description

A super-resolution face recognition method based on deep trust network

技术领域technical field

本发明涉及人脸识别技术领域，涉及到一种基于深度信赖网络的超分辨率人脸识别方法。The invention relates to the technical field of face recognition, and relates to a super-resolution face recognition method based on a deep trust network.

背景技术Background technique

人脸识别是一种重要的生物认证技术，是计算视觉和模式识别最重要的问题之一。近几十年来，研究人员提出了大量的方法，并已广泛用于视频监控等安全保障系统中。但是，由于距离和硬件条件等的限制，在大场景视频监控系统中拍摄的感兴趣人脸图像分辨率往往比较低，从而降低了人脸识别的性能。如何在低分辨率条件下提高识别效果，是目前人脸识别需要解决的问题。Face recognition is an important biometric authentication technology and one of the most important problems in computational vision and pattern recognition. In recent decades, researchers have proposed a large number of methods, which have been widely used in security systems such as video surveillance. However, due to the limitations of distance and hardware conditions, the resolution of the face images of interest captured in large-scene video surveillance systems is often relatively low, thereby reducing the performance of face recognition. How to improve the recognition effect under low-resolution conditions is a problem that needs to be solved in face recognition at present.

图像超分辨率(super-resolution,SR)是指利用某种算法从一幅或者一系列低分辨率(low resolution,LR)图像中获得一幅或者一系列高分辨率(high resolution,HR)图像的技术。因此，人脸图像超分辨率算法很自然地被作为提高低分辨率图像人脸识别效果的解决方案之一。申请号为：CN200810096054.7的专利：单帧图象超分辨方法，首先对图象进行分析，通过频率混叠参数判定是否采用单帧频域解混叠超分辨方法进行处理；然后通过傅立叶变换、频域解混叠算法及傅立叶反变换，丰富图象的纹理与细节，提高图象的清晰度、对比度和分辨率，并抑制振铃假象。此专利只适用于卫星遥感图象处理、医学图象和地震图象等图像的重建并非为了提高识别能力而设。这类方案将识别过程分解为人脸图像超分辨率重建和高分辨率人脸识别两步进行。然而人脸图像超分辨率重建的目标是尽可能地恢复高分辨率人脸图像的细节特征，以提高视觉效果，而影响人脸识别性能的特征可能既包括全局特征，又包括细节特征，两个步骤的目标不一致，导致最终的识别效果受到限制。基于以上原因，B.K.Gunturk和A.U.Batur在“Image Processing”（IEEE Trans.2003,vol.12,no.5,pp.597-606）发表的“Eigen-face-domainsuper-resolution for face recognition”等人提出了在特征域进行人脸超分辨率重建的方法，该方法超分辨率重建得到的特征可以直接用于人脸识别。该方法提供了一种很好的直接利用超分辨率算法进行人脸识别的框架，但计算复杂度较高，且该方法使用的概率模型对数据的一致性要求较高，当人脸姿态变化较大，算法的效果大幅度下降。B.Li,H.Chang,S.Shan and X.Chen在“Signal Processing Letters”（IEEE,2010,vol.17,no.1,pp.20-23）发表的“Low-resolution face recognition via coupled locality preservingmappings”提出了局部保持的耦合变换算法，利用局部保持对数据进行限定，耦合高低分辨率图像，从高低分辨率图像中抽取耦合特征。在姿态变化较大时，高低分辨率图像的局部保持性质差异较大，大大影响了其算法效果。专利申请号：CN200910207562.2的专利一种基于典型相关分析空间超分辨率的人脸识别方法，在典型相关分析变换得到的相关子空间内利用邻域重构获得测试低分辨率人脸图像对应的高分辨率人脸图像识别特征，最后利用此特征识别人脸。该方法在特征提取上仍然采用的是线性提取因子，其典型相关分析也是一种线性的变换方法，当存有较大姿态变化时，该方法性能大大降低。Image super-resolution (SR) refers to the use of a certain algorithm to obtain one or a series of high-resolution (high-resolution, HR) images from one or a series of low-resolution (low-resolution, LR) images. Technology. Therefore, the face image super-resolution algorithm is naturally regarded as one of the solutions to improve the face recognition effect of low-resolution images. The patent application number is: CN200810096054.7: Single-frame image super-resolution method. First, the image is analyzed, and the frequency aliasing parameter is used to determine whether to use the single-frame frequency-domain anti-aliasing super-resolution method for processing; and then through Fourier transform , frequency domain anti-aliasing algorithm and Fourier inverse transform, enrich the texture and details of the image, improve the clarity, contrast and resolution of the image, and suppress the ringing artifact. This patent is only applicable to satellite remote sensing image processing, reconstruction of medical images and seismic images, etc. It is not designed to improve the recognition ability. This type of scheme decomposes the recognition process into two steps: face image super-resolution reconstruction and high-resolution face recognition. However, the goal of face image super-resolution reconstruction is to restore the detailed features of high-resolution face images as much as possible to improve the visual effect, and the features that affect the performance of face recognition may include both global features and detailed features. The goals of each step are inconsistent, resulting in the limitation of the final recognition effect. Based on the above reasons, "Eigen-face-domain super-resolution for face recognition" published by B.K.Gunturk and A.U.Batur in "Image Processing" (IEEE Trans.2003, vol.12, no.5, pp.597-606) et al. A method of face super-resolution reconstruction in the feature domain is proposed, and the features obtained by super-resolution reconstruction of this method can be directly used for face recognition. This method provides a good framework for face recognition directly using the super-resolution algorithm, but the computational complexity is high, and the probability model used in this method has high requirements for data consistency. When the face pose changes Larger, the effect of the algorithm is greatly reduced. B.Li, H.Chang, S.Shan and X.Chen published "Low-resolution face recognition via coupled "locality preserving mappings" proposes a locality preserving coupling transformation algorithm, uses locality preservation to limit data, couples high and low resolution images, and extracts coupled features from high and low resolution images. When the pose changes greatly, the local preservation properties of high and low resolution images are quite different, which greatly affects the algorithm effect. Patent application number: CN200910207562.2, a face recognition method based on canonical correlation analysis space super-resolution, using neighborhood reconstruction in the correlation subspace obtained by canonical correlation analysis transformation to obtain the test low-resolution face image correspondence High-resolution face image recognition features, and finally use this feature to identify faces. This method still uses linear extraction factors in feature extraction, and its canonical correlation analysis is also a linear transformation method. When there is a large attitude change, the performance of this method is greatly reduced.

发明内容Contents of the invention

本发明克服上述现有技术的缺点，提出了一种基于深度信赖网络的超分辨率人脸识别方法。The present invention overcomes the above-mentioned shortcomings of the prior art, and proposes a super-resolution face recognition method based on a deep trust network.

为了达到上述目的，本发明采用的技术方案是：In order to achieve the above object, the technical scheme adopted in the present invention is:

本发明从认知的角度出发，认为相互对应的高低分辨率人脸图像存在着内在本质的关联。而以往的研究表明，采用线性近似的方法来表达这种内在的关联效果受到线性近似地制约。因此认为这种内在的关联是非线性的。鉴于人工神经网络在非线性分类问题上的出色表现，本发明采用神经网络算法来捕获姿态变化下相互对应的高低分辨率人脸图像的非线性关联。理论研究和神经生理学的研究表明，要构建一个智能的处理系统，需要构建深度的结构，如多层非线性处理单元构建的系统。对于构建深度网络，BP(back-propagation)算法是一种常用的神经网络算法。但是当网络的层数增加时，BP算法受到算法的局限，不能得到较好的结果。Hinton等人提出了快速地学习深度多层结构的概率模型的神经网络算法，将之命名为深度信赖网络（deep belief networks）。这种类型的神经网络不仅能作为分类器，而且可以表示非线性特征。基于此，本发明利用深度信赖网络（deep belief networks）来挖掘相互对应的高低分辨率人脸图像存在的共有的非线性结构。From the perspective of cognition, the present invention considers that there is an intrinsic relationship between the high and low resolution human face images corresponding to each other. However, previous studies have shown that the use of linear approximation to express this intrinsic correlation effect is restricted by linear approximation. Therefore, this internal correlation is considered to be non-linear. In view of the excellent performance of the artificial neural network on the nonlinear classification problem, the present invention adopts the neural network algorithm to capture the nonlinear association of the high and low resolution face images corresponding to each other under the attitude change. Theoretical studies and neurophysiological studies have shown that to build an intelligent processing system, it is necessary to build a deep structure, such as a system built by multi-layer nonlinear processing units. For building deep networks, BP (back-propagation) algorithm is a commonly used neural network algorithm. However, when the number of layers of the network increases, the BP algorithm is limited by the algorithm and cannot obtain better results. Hinton et al. proposed a neural network algorithm that quickly learns a probabilistic model of a deep multi-layer structure, and named it deep belief networks. This type of neural network can not only act as a classifier, but also represent nonlinear features. Based on this, the present invention utilizes deep belief networks (deep belief networks) to mine the shared non-linear structure existing in the corresponding high and low resolution face images.

附图说明Description of drawings

图1（a）是波尔兹曼机。Figure 1(a) is a Boltzmann machine.

图1（b）是受限的波尔兹曼机。Figure 1(b) is a restricted Boltzmann machine.

图2（a）是贪婪算法求得的受限的玻尔兹曼机。Figure 2(a) is the restricted Boltzmann machine obtained by the greedy algorithm.

图2（b）是深度信赖网络。Figure 2(b) is a deep trust network.

图2（c）是受限的波尔兹曼机构成的深度信赖网络。Figure 2(c) is a deep trust network composed of restricted Boltzmann machines.

图3（a）是UMIST图库中训练高分辨率56*46的人脸图像。Figure 3(a) is a training high-resolution 56*46 face image in the UMIST library.

图3（b）是训练低分辨率14*11的人脸图像。Figure 3(b) is a low-resolution 14*11 face image for training.

图3（c）是测试低分辨率14*11的人脸图像。Figure 3(c) is a test low-resolution 14*11 face image.

图4是超分辨率人脸识别算法示意图。Fig. 4 is a schematic diagram of a super-resolution face recognition algorithm.

具体实施方式Detailed ways

为使本发明的目的、技术方案及优点更加清楚明白，以下结合附图及具体实例，对本发明做进一步详细说明。这些实例仅仅是说明性的，而并非对本发明的限制。In order to make the object, technical solution and advantages of the present invention clearer, the present invention will be further described in detail below in conjunction with the accompanying drawings and specific examples. These examples are illustrative only and not limiting of the invention.

本发明提出了一种基于深度信赖网络的超分辨率人脸识别方法，该方法可以包括：The present invention proposes a super-resolution face recognition method based on deep trust network, which method may include:

a.受限的波尔兹曼机。受限的波尔兹曼机是一种马尔科夫随机场或者一种双层图结构，一种特殊结构的波尔兹曼机。如图1所示，图（a）为一般的波尔兹曼机，波尔兹曼机为一种满秩的双层图结构，可将下层称为可视层，上层称为隐层。（b）为受限的波尔兹曼机，受限的波尔兹曼机与一般的波尔兹曼机相比，不允许可视层各点之间或者是隐层各点之间存在关联。a. Restricted Boltzmann machine. Restricted Boltzmann machine is a Markov random field or a double-layer graph structure, a special structure of Boltzmann machine. As shown in Figure 1, the picture (a) is a general Boltzmann machine, and the Boltzmann machine is a full-rank two-layer graph structure. The lower layer can be called the visible layer, and the upper layer can be called the hidden layer. (b) is a restricted Boltzmann machine. Compared with the general Boltzmann machine, the restricted Boltzmann machine does not allow the existence of associated.

受限的波尔兹曼机是一种特殊的神经网络模型，具有对称的连接权系数。网络由可视单元v∈{0,1}^D和隐层单元h∈{0,1}^F构成。可视单元由输入、输出部分组成。每个单元节点只取1或0两种状态。1代表接通或接受，0表示断开或拒绝。当神经元的输入加权和发生变化时，神经元的状态随之更新。各单元之间状态的更新是异步的。可用概率来描述。状态{v,h}的能量方程可以定义为：Restricted Boltzmann machine is a special neural network model with symmetric connection weight coefficients. The network consists of visible units v∈{0,1} ^D and hidden layer units h∈{0,1} ^F . Visual unit consists of input and output parts. Each unit node only takes two states of 1 or 0. 1 means connected or accepted, 0 means disconnected or rejected. When the weighted sum of inputs to a neuron changes, the state of the neuron is updated accordingly. State updates between units are asynchronous. It can be described by probability. The energy equation for state {v,h} can be defined as:

E(v,h;θ)=-v^TWh-b^Tv-a^Th （1）E(v,h;θ)=-v ^T Wh-b ^T va ^T h (1)

其中θ={W,a,b}为参数，W可视层和隐层之间对称的连接权系数，a和b为基矩阵。可见向量与隐层向量的联合分布矩阵为：Where θ={W,a,b} is a parameter, W is a symmetrical connection weight coefficient between the visible layer and the hidden layer, and a and b are the base matrices. The joint distribution matrix of the visible vector and the hidden layer vector is:

$p p ((v v,, h h;; θ θ)) = = \frac{11}{z z ((θ θ))} exp exp ((- - E E. ((v v,, h h;; θ θ)))) - - - - - - ((22))$

$z z ((θ θ)) = = \underset{v v}{Σ Σ} \underset{h h}{Σ Σ} exp exp ((- - E E. ((v v,, h h;; θ θ)))) - - - - - - ((33))$

模型分配可见向量的概率为：The probability that the model assigns a visible vector is:

$p p ((v v,, h h;; θ θ)) = = \frac{11}{z z ((θ θ))} \underset{h h}{Σ Σ} exp exp ((- - E E. ((v v,, h h;; θ θ)))) - - - - - - ((44))$

可见向量和隐层向量的条件概率分布为：The conditional probability distribution of visible vector and hidden layer vector is:

$p p (({h h}_{j j} = = 11 | | v v)) = = g g ((\underset{i i}{Σ Σ} {W W}_{ij ij} {v v}_{i i} + + {a a}_{j j})) - - - - - - ((55))$

$p p (({v v}_{i i} = = 11 | | h h)) = = g g ((\underset{j j}{Σ Σ} {W W}_{ij ij} {h h}_{j j} + + {b b}_{i i})) - - - - - - ((66))$

其中g(x)=1/(1+exp(-x))。对各参数求导可得其指数似然概率：where g(x)=1/(1+exp(-x)). The exponential likelihood probability can be obtained by deriving each parameter:

$\frac{&PartialD; &PartialD; log log p p ((v v;; θ θ))}{&PartialD; &PartialD; W W} = = α α (({E E.}_{date date} [[{vh vh}^{T T}]] - - {E E.}_{mode mode l l} [[{vh vh}^{T T}]])) - - - - - - ((77))$

$\frac{&PartialD; &PartialD; log log p p ((v v;; θ θ))}{&PartialD; &PartialD; a a} = = α α (({E E.}_{date date} [[h h]] - - {E E.}_{mode mode l l} [[h h]])) - - - - - - ((88))$

$\frac{&PartialD; &PartialD; log log p p ((v v;; θ θ))}{&PartialD; &PartialD; b b} = = α α (({E E.}_{date date} [[v v]] - - {E E.}_{mode mode l l} [[v v]])) - - - - - - ((99))$

其中α为学习率。E_data[.]为数据完整分布p_data(v,h;θ)=p(h|v;θ)p_data(v)的期望，其中p_data(v)为数据的先验知识。E_model[.]为公式（2）代表的模型期望。数据期望和模型期望是很难获得的，可以采用吉布斯采样获得上述期望的近似值。在实际中，通过马尔科夫链估计模型期望，通过变分法估计数据期望。where α is the learning rate. E _data [.] is the expectation of the complete data distribution p _data (v,h;θ)=p(h|v;θ)p _data (v), where p _data (v) is the prior knowledge of the data. E _model [.] is the model expectation represented by formula (2). Data expectations and model expectations are difficult to obtain, and Gibbs sampling can be used to obtain approximate values of the above expectations. In practice, model expectations are estimated through Markov chains, and data expectations are estimated through variational methods.

深度信赖网络是一个有多层隐层的概率模型，每一层从前一层的隐含单元捕获高度相关的关联。图2为一个深度信赖网络的示意图。相邻两层可分解为一个单独的受限的波尔兹曼机。A deep trust network is a probabilistic model with multiple hidden layers, each layer capturing highly correlated associations from hidden units in the previous layer. Figure 2 is a schematic diagram of a deep trust network. Two adjacent layers can be decomposed into a single restricted Boltzmann machine.

b.训练深度信赖网络b. Training deep trust network

深度信赖网络的全局优化比较困难。为了取得较好的训练效果，可以采取逐层贪婪算法，每次只训练相邻两层之间的参数和隐层数据，逐层计算获得最终的深度信赖网络。通过贪婪算法获得的深度信赖网络根据最终的感兴趣的判别准则进行微调就可以获得最终深度信赖网络。Global optimization of deep trust networks is difficult. In order to obtain a better training effect, a layer-by-layer greedy algorithm can be adopted, and only the parameters and hidden layer data between two adjacent layers are trained each time, and the final deep trust network is obtained by layer-by-layer calculation. The deep trust network obtained by the greedy algorithm can be fine-tuned according to the final criterion of interest to obtain the final deep trust network.

（1）初始化(1) Initialization

如图2所示，将深度信赖网络分解成由相邻两层构成的一系列受限的玻尔兹曼机，逐层训练参数，初始化深度信赖网络：As shown in Figure 2, the deep trust network is decomposed into a series of restricted Boltzmann machines consisting of two adjacent layers, the parameters are trained layer by layer, and the deep trust network is initialized:

首先将经验数据v作为输入，训练第一层受限波尔兹曼机的权值系数矩阵W¹；接着将W¹固定，通过p(h¹|v)=p(h¹|v,W¹)，训练出第一层受限波尔兹曼机的隐层向量h¹；将h¹作为第二层受限波尔兹曼机的输入，训练第二层受限波尔兹曼机的权值系数矩阵W²；递归地计算出每一层的隐含单元向量和权值系数矩阵。Firstly, the empirical data v is used as input to train the weight coefficient matrix W ¹ of the first-layer restricted Boltzmann machine; then W ¹ is fixed, by p(h ¹ |v)=p(h ¹ |v,W ¹ ), train the hidden layer vector h ¹ of the first-layer restricted Boltzmann machine; use h ¹ as the input of the second-layer restricted Boltzmann machine, and train the second-layer restricted Boltzmann machine The weight coefficient matrix W ² ; recursively calculate the hidden unit vector and weight coefficient matrix of each layer.

（2）微调(2) fine-tuning

为达到分类目的，本发明最终的微调在最后一层网络再添加一层逻辑回归层，并采用梯度下降法，保证目标分类误差最小，训练整个网络。In order to achieve the purpose of classification, the final fine-tuning of the present invention adds a layer of logistic regression to the last layer of the network, and adopts the gradient descent method to ensure the minimum classification error of the target and train the entire network.

本发明具体的实现步骤如下：Concrete implementation steps of the present invention are as follows:

（1）首先，对低分辨率图像进行最近邻插值或者双线性插值，使得高低分辨率图像的维度一致；(1) First, the nearest neighbor interpolation or bilinear interpolation is performed on the low-resolution image, so that the dimensions of the high-resolution and low-resolution images are consistent;

（2）接着，如图3所示，将维度一致的带有姿态差异的高低分辨率人脸图像作为深度信赖网络的可视向量输入到网络中，深度信赖网络由受限的波尔兹曼机构成；(2) Next, as shown in Figure 3, high- and low-resolution face images with consistent dimensions and pose differences are input into the network as the visual vector of the deep trust network. The deep trust network consists of a restricted Boltzmann Institutional composition;

（3）然后，训练深度信赖网络。训练深度信赖网络主要包含初始化和微调两步。初始化时，通过相邻层之间的重建误差控制，采用贪婪算法逐层计算深度信赖网络参数。微调时，采用标准的BP（back-propagation）算法，保证分类误差最小，训练整个深度信赖网络；(3) Then, train a deep trust network. Training a deep trust network mainly includes two steps of initialization and fine-tuning. During initialization, through the reconstruction error control between adjacent layers, the greedy algorithm is used to calculate the parameters of the deep trust network layer by layer. When fine-tuning, the standard BP (back-propagation) algorithm is used to ensure the minimum classification error and train the entire deep trust network;

（4）最后，将测试低分辨率图像最近邻插值或者双线性插值到高分辨率图像大小，输入到深度信赖网络，由深度信赖网络给出最终的识别结果。(4) Finally, the nearest neighbor interpolation or bilinear interpolation of the test low-resolution image to the size of the high-resolution image is input to the deep trust network, and the final recognition result is given by the deep trust network.

图3为用于实验的UMIST图库中一个人物的人脸图像。其中图为（a）训练高分辨率人脸图像，图像大小为56*46，图（b）为训练低分辨率人脸图像，图像大小为14*11，图（c）为测试低分辨率人脸图像，图像大小为14*11。Figure 3 is a face image of a person in the UMIST library used for experiments. The picture is (a) training high-resolution face image, the image size is 56*46, picture (b) is training low-resolution face image, the image size is 14*11, picture (c) is testing low-resolution Face image, the image size is 14*11.

图4为UMIST图库采用深度信赖网络进行超分辨率人脸识别的示意图。各个隐层所含单元数由实际效果决定。Figure 4 is a schematic diagram of UMIST library using deep trust network for super-resolution face recognition. The number of units contained in each hidden layer is determined by the actual effect.

从上述描述应该理解，在不脱离本发明精神的情况下，可以对本发明各实施方式进行修改和变更。本说明书中的描述仅仅是用于说明性的，而不应被认为是限制性的。It should be understood from the above description that modifications and changes can be made to the various embodiments of the present invention without departing from the spirit of the present invention. The descriptions in this specification are for illustration only and should not be considered as limiting.

Claims

1. trust the super-resolution face recognition method of network based on the degree of depth for one kind, it is characterized in that comprising following steps:

1) low-resolution image is carried out arest neighbors interpolation, bilinear interpolation or bicubic interpolation, so that the dimension of high low-resolution image is consistent;

2) the high low resolution facial image gray scale normalizing with attitude difference that dimension is consistent arrives between (0,1), and is input in the network as the visual vector v of degree of depth trust network, and degree of depth trust network is made of the limited Boltzmann machine of multilayer; Described limited Boltzmann machine is a kind of special neural network model, has symmetrical link weight coefficients, and network is by visual element v ∈ { 0,1} ^DWith Hidden unit h ∈ { 0,1} ^FConsist of;

3) then, the training degree of depth is trusted network;

4) will be input to through the test low-resolution image of arest neighbors interpolation, bilinear interpolation or bicubic interpolation the degree of depth and trust network, and trust network by the degree of depth and provide final recognition result.

2. super-resolution face recognition method as claimed in claim 1, it is characterized in that: described step 2) refer to: high its resolution of low resolution facial image with attitude difference is h * w, its expansion is become the vector that delegation's length is h * w, and with its gray-scale intensity normalizing to (0,1).

3. super-resolution face recognition method as claimed in claim 1 or 2 is characterized in that: the training degree of depth in the described step 3) is trusted network and is comprised following steps:

The degree of depth is trusted network resolve into a series of limited Boltzmann machine that is consisted of by adjacent two layers, successively training parameter;

1) the initialization degree of depth is trusted network: at first with empirical data v as input, the weights matrix of coefficients W of the limited Boltzmann machine of training ground floor ¹Then with W ¹Fixing, by p (h ¹| v)=p (h ¹| v, W ¹), train the hidden layer vector h of the limited Boltzmann machine of ground floor ¹With h ¹As the input of the limited Boltzmann machine of the second layer, the weights matrix of coefficients W of the limited Boltzmann machine of the training second layer ²Recursively calculate the implicit unit vector sum weights matrix of coefficients of every one deck;

2) fine setting: for reaching the classification purpose, in the end a layer network adds first level logical recurrence layer again, and adopts gradient descent method to train whole network.

4. super-resolution face recognition method as claimed in claim 3 is characterized in that: the described initialization degree of depth is trusted network, calculates the following characteristics that comprises of weights and Hidden unit:

By empirical data estimation model parameter or Hidden unit state: state v, the energy equation of h} is defined as:

E(v,h;θ)=-v ^TWh-b ^Tv-a ^Th （1）

Wherein θ=W, a, b} are parameter, symmetrical link weight coefficients between W visual layers and the hidden layer, a and b are basis matrix, the joint distribution matrix of visual vector and hidden layer vector is:

p (v, h; θ) = \frac{1}{z (θ)} \exp (- E (v, h; θ)) - - - (2)

If E _Data[.] is the data integrity distribution p _Data(v, h; θ)=p (h|v; θ) p _Data(v) expectation, wherein p _Data(v) be the priori of data, E _Model[.] is the model expectation of formula (2) representative, then tries to achieve optimum parameter θ={ W, a, b} or corresponding hidden layer state vector h by formula (3 ~ 5).

\frac{&PartialD; \log p (v; θ)}{&PartialD; W} = α (E_{date} [{vh}^{T}] - E_{mode l} [{vh}^{T}]) - - - (3)

\frac{&PartialD; \log p (v; θ)}{&PartialD; a} = α (E_{date} [h] - E_{mode l} [h]) - - - (4)

\frac{&PartialD; \log p (v; θ)}{&PartialD; b} = α (E_{date} [v] - E_{mode l} [v]) - - - (5)

Wherein α is learning rate.