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

Abstract

The invention discloses a super-resolution face recognition method based on deep belief networks, and relates to the technical field of face recognition. From a cognitive perspective, it is believed that an intrinsic relation exists between mutually corresponding face images differing in resolutions. But previous studies show that the method for expressing the intrinsic relation by linear approximation is restricted by linear approximation. Therefore, it is believed that the intrinsic relation is nonlinear. In view of outstanding performances of an artificial neural network on the nonlinear classification problem, a neural network algorithm is adopted to capture the nonlinear relation of the mutually corresponding face images differing in resolutions under the condition of posture change. Both theoretical research and neurophysiological research show that a deep structure, such as a system constructed by multiple layers of nonlinear processing units, should be constructed to build an intelligent processing system. According to the face recognition method, deep belief networks are adopted to extract a common nonlinear structure shared by mutually corresponding face images differing in resolutions.

Description

A kind of super-resolution face recognition method of trusting network based on the degree of depth

Technical field

The present invention relates to the face recognition technology field, relate to a kind of super-resolution face recognition method of trusting network based on the degree of depth.

Background technology

Recognition of face is a kind of important biological identification technology, is one of computation vision and pattern-recognition sixty-four dollar question.In recent decades, the researchist has proposed a large amount of methods, and has been widely used in the safe-guard systems such as video monitoring.But because the restriction of distance and hardware condition etc., the facial image resolution interested of taking in the large scene video monitoring system is often lower, thereby has reduced the performance of recognition of face.How to improve recognition effect under the low resolution condition, be the problem that present recognition of face need to solve.

Image super-resolution (super-resolution, SR) refer to utilize certain algorithm from a width of cloth or a series of low resolution (low resolution, LR) technology of acquisition one width of cloth or a series of high resolving power (high resolution, HR) image in the image.Therefore, the face image super-resolution algorithm is used as one of solution that improves low-resolution image recognition of face effect very naturally.Application number is: the patent of CN200810096054.7: method for super resolution of single-frame images, at first image is analyzed, and whether adopt single frames Frequency Domain Solution aliasing ultra-resolution method to process by the frequency alias parameter decision; Then by Fourier transform, Frequency Domain Solution aliasing algorithm and inverse fourier transform, texture and the details of abundant image improve visual sharpness, contrast and resolution, and the suppressed ringing illusion.The reconstruction that this patent is only applicable to the images such as processing of remotely sensed satellite image, medical images and earthquake vision is not to establish in order to improve recognition capability.Identifying is decomposed into face image super-resolution rebuilding to this class scheme and two steps of high-resolution human face identification carry out.Yet the target of face image super-resolution rebuilding is to recover as much as possible the minutia of high-resolution human face image, to improve visual effect, and the feature that affects the recognition of face performance may both comprise global characteristics, comprise again minutia, the target of two steps is inconsistent, causes final recognition effect to be restricted.Based on above reason, B.K.Gunturk and A.U.Batur are at " Image Processing " (IEEE Trans.2003, vol.12, no.5, pp.597-606) people such as " the Eigen-face-domainsuper-resolution for face recognition " that deliver has proposed to carry out the method that human face super-resolution is rebuild at property field, and the feature that the method super-resolution rebuilding obtains can be directly used in recognition of face.The method provides a kind of and has well directly utilized super-resolution algorithms to carry out the framework of recognition of face, but computation complexity is higher, and the probability model that the method is used is higher to the coherence request of data, and when human face posture changes greatly, the effect of algorithm significantly descends.B.Li, H.Chang, S.Shan and X.Chen is at " Signal Processing Letters " (IEEE, 2010, vol.17, no.1, pp.20-23) " the Low-resolution face recognition via coupled locality preservingmappings " that delivers proposed the coupling transform algorithm of local maintenance, utilize local the maintenance that data are limited, the high low-resolution image that is coupled extracts coupling feature from high low-resolution image.When attitude changed greatly, the local retention properties of high low-resolution image differed greatly, and has greatly affected its algorithm effect.A kind of face identification method based on typical correlation analysis spatial super-resolution of the patent of number of patent application: CN200910207562.2, in the correlator space that the canonical correlation analysis conversion obtains, utilize neighborhood reconstruct to obtain high-resolution human face image recognition feature corresponding to test low resolution facial image, utilize at last this feature identification people face.What the method still adopted in feature extraction is linear extraction factor, and its canonical correlation analysis also is a kind of transform method of linearity, and when having larger attitude variation, the method performance reduces greatly.

Summary of the invention

The present invention overcomes the shortcoming of above-mentioned prior art, has proposed a kind of super-resolution face recognition method of trusting network based on the degree of depth.

In order to achieve the above object, the technical solution used in the present invention is:

The present invention thinks that from the angle of cognition the high low resolution facial image of mutual correspondence exists the association of inward nature.And studies show that in the past adopts the method for linear-apporximation to express the interrelating effect restriction of this inherence with being subject to linear-apporximation.Therefore think that the association of this inherence is nonlinear.In view of the outstanding performance of artificial neural network on the Nonlinear Classification problem, the present invention adopts neural network algorithm to catch the non-linear correlation that attitude changes the high low resolution facial image of lower mutual correspondence.Studies show that of theoretical research and neuro-physiology will make up the disposal system of an intelligence, needs to make up the structure of the degree of depth, the system that makes up such as the multilayered nonlinear processing unit.For making up degree of depth network, BP (back-propagation) algorithm is a kind of neural network algorithm commonly used.But when the number of plies of network increased, the BP algorithm was subject to the limitation of algorithm, can not obtain preferably result.The people such as Hinton have proposed to learn rapidly the neural network algorithm of the probability model of degree of depth sandwich construction, and it called after degree of depth is trusted network (deep belief networks).Such neural network can not only be as sorter, and can represent nonlinear characteristic.Based on this, the present invention utilizes the degree of depth to trust the total nonlinear organization that network (deep belief networks) excavates the high low resolution facial image existence of mutual correspondence.

Description of drawings

Fig. 1 (a) is Boltzmann machine.

Fig. 1 (b) is limited Boltzmann machine.

Fig. 2 (a) is the limited Boltzmann machine that greedy algorithm is tried to achieve.

Fig. 2 (b) is that the degree of depth is trusted network.

Fig. 2 (c) is that the degree of depth that limited Boltzmann machine consists of is trusted network.

Fig. 3 (a) is the facial image of training high resolving power 56*46 in the UMIST picture library.

Fig. 3 (b) is the facial image of training low resolution 14*11.

Fig. 3 (c) is the facial image of test low resolution 14*11.

Fig. 4 is super-resolution face recognition algorithms synoptic diagram.

Embodiment

For making purpose of the present invention, technical scheme and advantage clearer, below in conjunction with accompanying drawing and instantiation, the present invention is described in further details.These examples only are illustrative, and are not limitation of the present invention.

The present invention proposes a kind of super-resolution face recognition method based on degree of depth trust network, the method can comprise:

A. limited Boltzmann machine.Limited Boltzmann machine is a kind of markov random file or a kind of double-deck graph structure, a kind of Boltzmann machine of special construction.As shown in Figure 1, figure (a) is general Boltzmann machine, and Boltzmann machine is a kind of double-deck graph structure of full rank, lower floor can be called visual layers, and the upper strata is called hidden layer.(b) be limited Boltzmann machine, limited Boltzmann machine is compared with general Boltzmann machine, does not allow between the visual layers each point or existence association between the hidden layer each point.

Limited Boltzmann machine is a kind of special neural network model, has symmetrical link weight coefficients.Network is by visual element v ∈ { 0,1} ^DWith Hidden unit h ∈ { 0,1} ^FConsist of.Visual element is comprised of input, output.Each cell node is only got 1 or 0 two states.1 representative is connected or is accepted, and 0 expression disconnects or refusal.When neuronic weighted input with when changing, neuronic state upgrades thereupon.The renewal of state is asynchronous between each unit.Usable probability is described.State v, the energy equation of h} can be 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.As seen vector with the joint distribution matrix of hidden layer vector is:

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

$z\left(\mathrm{\θ}\right)=\underset{v}{\mathrm{\Σ}}\underset{h}{\mathrm{\Σ}}\exp (-E(v,h;\mathrm{\θ}))---\left(3\right)$

Model distributes the probability of visible vector to be:

$p(v,h;\mathrm{\θ})=\frac{1}{z\left(\mathrm{\θ}\right)}\underset{h}{\mathrm{\Σ}}\exp (-E(v,h;\mathrm{\θ}))---\left(4\right)$

As seen the conditional probability distribution of vector sum hidden layer vector is:

$p(h_j=1|v)=g(\underset{i}{\mathrm{\Σ}}{W}_{\mathrm{ij}}{v}_i+a_j)---\left(5\right)$

$p(v_i=1|h)=g(\underset{j}{\mathrm{\Σ}}{W}_{\mathrm{ij}}{h}_j+b_i)---\left(6\right)$

G (x)=1/ (1+exp (x)) wherein.Can get its index likelihood probability to each parameter differentiate:

$\frac{\∂\log p(v;\mathrm{\θ})}{\∂W}=\mathrm{\α}\left(E_{\mathrm{date}}\right[{\mathrm{vh}}^T]-E_{\mathrm{mode}l}[{\mathrm{vh}}^T\left]\right)---\left(7\right)$

$\frac{\∂\log p(v;\mathrm{\θ})}{\∂a}=\mathrm{\α}\left(E_{\mathrm{date}}\right[h]-E_{\mathrm{mode}l}[h\left]\right)---\left(8\right)$

$\frac{\∂\log p(v;\mathrm{\θ})}{\∂b}=\mathrm{\α}\left(E_{\mathrm{date}}\right[v]-E_{\mathrm{mode}l}[v\left]\right)---\left(9\right)$

Wherein α is learning rate.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.Data expectation and model are contemplated to be very unobtainable, can adopt gibbs sampler to obtain the approximate value of above-mentioned expectation.In practice, by the expectation of Markov chain estimation model, expect by variational method data estimator.

It is a probability model that the multilayer hidden layer is arranged that the degree of depth is trusted network, and every one deck is the association of the implicit elements capture height correlation of one deck in the past.Fig. 2 is the synoptic diagram that a degree of depth is trusted network.Adjacent two layers can be decomposed into an independent limited Boltzmann machine.

B. train the degree of depth to trust network

The global optimization that the degree of depth is trusted network is relatively more difficult.In order to obtain preferably training effect, can take successively greedy algorithm, only train parameter and hidden layer data between the adjacent two layers at every turn, successively calculate and obtain final degree of depth trust network.The degree of depth that obtains by greedy algorithm is trusted network based final interested criterion and is finely tuned and just can obtain ultimate depth trust network.

(1) initialization

As shown in Figure 2, the degree of depth is trusted network resolve into a series of limited Boltzmann machine that is consisted of by adjacent two layers, training parameter successively, the initialization degree of depth is trusted network:

At first with empirical data v as input, the weights matrix of coefficients W of training ground floor limited Boltzmann machine 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, the final fine setting of a present invention in the end layer network is added first level logical again and is returned layer, and adopts gradient descent method, guarantees that the target classification error is minimum, trains whole network.

The concrete performing step of the present invention is as follows:

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

(2) then, as shown in Figure 3, the visual vector that the high low resolution facial image with attitude difference that dimension is consistent is trusted network as the degree of depth is input in the network, and the degree of depth is trusted network and is made of limited Boltzmann machine;

(3) then, the training degree of depth is trusted network.The training degree of depth is trusted network and is mainly comprised initialization and two steps of fine setting.During initialization, by the control of the reconstruction error between the adjacent layer, adopt successively compute depth trust of greedy algorithm network parameter.During fine setting, adopt the BP(back-propagation of standard) algorithm, guarantee that error in classification is minimum, the training entire depth is trusted network;

(4) last, will test low-resolution image arest neighbors interpolation or bilinear interpolation to the high-definition picture size, be input to the degree of depth and trust network, trust network by the degree of depth and provide final recognition result.

Fig. 3 is the facial image for a personage of UMIST picture library who tests.Wherein figure is (a) training high-resolution human face image, and the image size is 56*46, and figure (b) is training low resolution facial image, and the image size is 14*11, and figure (c) is test low resolution facial image, and the image size is 14*11.

Fig. 4 is that the UMIST picture library adopts the degree of depth to trust the synoptic diagram that network carries out the super-resolution recognition of face.The contained unit number of each hidden layer is determined by actual effect.

Should be appreciated that from foregoing description, in the situation that does not break away from spirit of the present invention, can make amendment and change each embodiment of the present invention.Description in this instructions is only used for illustrative, and should not be considered to restrictive.

Claims (4)

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:

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

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).

$\frac{\∂\log p(v;\mathrm{\θ})}{\∂W}=\mathrm{\α}\left(E_{\mathrm{date}}\right[{\mathrm{vh}}^T]-E_{\mathrm{mode}l}[{\mathrm{vh}}^T\left]\right)---\left(3\right)$

$\frac{\∂\log p(v;\mathrm{\θ})}{\∂a}=\mathrm{\α}\left(E_{\mathrm{date}}\right[h]-E_{\mathrm{mode}l}[h\left]\right)---\left(4\right)$

$\frac{\∂\log p(v;\mathrm{\θ})}{\∂b}=\mathrm{\α}\left(E_{\mathrm{date}}\right[v]-E_{\mathrm{mode}l}[v\left]\right)---\left(5\right)$

Wherein α is learning rate.

Images