CN106682653A - KNLDA-based RBF neural network face recognition method - Google Patents

Abstract

The invention relates to the technical field of face recognition. The face recognition method includes steps of combining advantages of null space linear distinguishing analysis (NLDA) and kernel function. The null space linear distinguishing analysis (NLDA) extracts identifying characteristics of a sample from a null space of a dispersion matrix of total classification of a training sample, thereby overcoming problems of small samples and improving the recognition rate; however, the NLDA is still an extracting method of the linear characteristics, and non-linear characteristics of the sample cannot be effectively extracted. Through the non-linear mapping, the method maps the input space sample to the high-dimensional characteristics space; the linear characteristic extracting algorithm to a high-dimensional characteristics space, thereby effectively extracting the non-linear characteristics of the sample. We use the RBP neutral network to identify the face image acquired through the characteristics extraction. The study indicates the recognition efficiency can be improved by comparing with a traditional classification method based on Europe type distance and other traditional measurements. The technical scheme provided by the invention can be well applied to the actual life, the recognition rate is higher, and the robustness is better.

Description

A kind of RBF neural face identification method based on KNLDA

Technical field

The present invention relates to a kind of biometric discrimination method, and in particular to one kind is based on core kernel linear discriminant analysis (KNLDA) face identification method of RBF neural, belongs to computer vision and area of pattern recognition.

Background technology

Recognition of face is an emphasis research topic of current pattern recognition.Over nearly 20 years, the research of face recognition technology Have become a study hotspot, it is simple, directly, be easy to be easily accepted by a user.At present, face recognition technology is extensively applied In fields such as certificate verification, intelligent monitoring, criminal investigation and case detections, but because recognition of face is that non-rigid, illumination variation, posture become Change, change of age, the problems such as block, these produce impact to face recognition algorithms, and its degree is different.

Because the dimensional comparison of facial image is high, the usual way of conventional method is that facial image is carried out into dimensionality reduction to extract special Face is levied, then is contrasted.It is dimensionality reduction that feature extraction is the committed step its main purpose of recognition of face, and more classical algorithm has , compared to the PCA for the purpose of reconstruct, the LDA for the purpose of classification exists for principal component analysiss (PCA) and linear discriminant analysiss (LDA) During process recognition of face problem more effectively.But small sample problem is frequently encountered in recognition of face makes matrix of taking a walk in class Sw is unusual, and for the problem, researcher proposes many algorithms in succession, it is proposed that kernel linear discriminant analysiss (NLDA) are simultaneously demonstrate,proved Real General optimal discriminant vectors should be located in the kernel of Sw.NLDA methods are substantially still a kind of linear method, are known in face In other problem, due to differences such as illumination, attitude, expressions, facial image distribution is caused to be nonlinear and complicated, so linearly Algorithm is when the image recognition tasks such as face are completed, it is impossible to achieve satisfactory results.By nonlinear mapping, kernel method will Input space sample is mapped to high-dimensional feature space, Linear feature extraction algorithm is utilized in high-dimensional feature space, so as to effectively carry Sampling nonlinear characteristic originally.NLDA is combined with kernel method, core kernel linear discriminant analysis (KNLDA) is obtained, is led KNLDA algorithms are gone out, by introducing kernel function, have obtained low-dimensional matrix, effectively prevent and directly calculate complicated nonlinear mapping Function, solves the problems, such as the dimension disaster of higher-dimension within class scatter matrix.

RBF (Radial Basis Function, RBF) neutral net be in last century late nineteen eighties by A kind of neutral net that J.moody and C.Darken is proposed, it simulates adjustment in human brain, mutually covers the nerve of acceptance region Network structure, can be a kind of partial approximation network with any progress arbitrary continuation function after all, and the structure of RBF networks being capable of root It is adjusted according to particular problem, it is fast with pace of learning, the characteristics of local minimum will not be absorbed in, system is widely used at present Identification, the multiple fields such as function approximation, signal processing and control.

In real life, facial image is by illuminating, posture, at the age, the change such as block and cause recognition of face difficulty.Institute With.In daily life, face recognition technology is satisfactory not enough, is that this carries out for a long time in-depth study, and recent years, face is known Other technology has obtained very big progress, but still can not all reach satisfied requirement.

The content of the invention

Recognition of face is abnormal serious by the interference of illumination, attitude, a series of natural causes of shelter, in order to improve identification Rate, enhancing robustness, the invention provides a kind of RBF neural of the linear discriminant analysis (KNLDA) based on core kernel Face identification method, the method can preferably be embodied the nonlinear characteristic of face, be carried out more accurately using RBF neural Identification.

The present invention solves above-mentioned technical problem by following technological means：

A kind of face identification method of the RBF neural based on KNLDA, its step is as follows：First, facial image is read Go forward side by side pedestrian's face region detection, then, facial image strengthens pretreatment, best projection matrix is chosen by KNLDA then, to people Face image carries out feature extraction, and last Training RBF Neural Network realizes recognition of face.

Traditional LDA methods are changed greatly, exist under the natural environment of partial occlusion for illumination relatively strong, human face posture, Discrimination can drastically decline, in order to improve the robustness and discrimination of method, herein using the RBF neural based on KNLDA Face identification method, its step is as follows：

(1) input space is mapped to into higher dimensional space by Kernel-Based Methods；

(2) sample of higher dimensional space is carried out into kernel linear discriminant analysis, can effectively solving small sample problem, and make Inter- object distance is minimum, and between class distance is maximum；

(3) best projection matrix is obtained, dimensionality reduction matrix is tried to achieve, with this sample training RBF neural；

(4) trained using Orthogonal Least Square learning algorithm, by orthogonalization method, chosen in the input space to drop The larger vector of low error contribution is used as network center；

(5) recognition of face is realized using RBF neural.

RBF neural has advantages below：

1st, the characteristic most preferably approached with the overall situation.

2nd, the input of network and output mapping function are stronger, and theory confirms that compared with other feedforward networks, RBF networks are Complete the optimum feedforward network of mapping function.

3rd, linear relationship is presented between connection weight and output in network.

4th, stronger classification capacity.

5th, because neuron carries out local modulation, the learning process fast convergence rate of RBF neural.

The present invention is changed greatly for intensity of illumination, human face posture, face presence is blocked and causes discrimination seriously to reduce, and is carried A kind of face identification method of the RBF neural based on KNLDA is gone out, projection can farthest have been optimized by the method Matrix, so that object function reaches maximum, and the more conventional two-dimensional principal component analysis method method of discrimination is compared, identification Rate is higher, and robustness is more preferable.

Description of the drawings

Fig. 1 is the face identification method flow chart based on the RBF neural of KNLDA

Fig. 2 is based on the colour of skin and AdaBoost method for detecting human face flow charts

Fig. 3 is the structural representation of RBF neural

Fig. 4 is software initial interface figure

Fig. 5 is sample registered surface chart

Fig. 6 is training sample exemplary plot

Fig. 7 is running software initialization surface chart

Fig. 8 is training surface chart

Fig. 9 is the recognition effect figure under complicated case

Specific embodiment

The thinking of the present invention is in illumination, attitude, a series of natural causes of shelter for existing face identification method Interference under, discrimination is greatly lowered, the problem that robustness also weakens, it is proposed that a seed nucleus kernel linear discriminant analysis (KNLDA) face identification method of RBF neural, by introducing kernel function, is mapped to high dimensional feature empty by the input space Between, the pattern of lower dimensional space linearly inseparable may then realize linear separability by nonlinear mapping to high-dimensional feature space, grind Study carefully and show, compared with the classification based on euclidean distance metric, using RBF neural classification method recognition of face system can be improved System.Improve face identification method larger in illumination, attitudes vibration, there is the discrimination under partial occlusion, strengthen robustness.

The face identification method of the present invention, its flow chart is as shown in Figure 1：Specifically according to following steps:

Step 1, Face datection

It is a considerably complicated feature in view of face, needs to consider more factor in Face datection, based on skin The detection method of color has the stronger suitability to the change such as human face expression, attitude, however, this method false drop rate is higher, is based on Although the method for detecting human face of AdaBoost possesses relatively low false drop rate, but its detection speed is slower.Consider, in order to carry The performance of high method for detecting human face, adopts herein by AdaBoost methods in combination with complexion model, so as to complete Face datection. May be summarized to be based on the thought of colour of skin characteristic and AdaBoost method for detecting human face：First, using human body complexion Clustering features Set up based on the statistical model of the colour of skin under YCbCr color spaces, human face region is filtered out from altimetric image to be checked, use then The cascade classifier that improved AdaBoost is trained verifies to possible human face region, finally determines the face in image Position, the method for detecting human face flow chart of the present invention is as shown in Figure 2.

Step 2, Image semantic classification

It is such as warm by external environment condition during carrying out shooting video image and transmission using image capture device Degree, illumination and the impact of equipment factor itself, the picture quality for getting can be reduced.Accordingly, it would be desirable to the image for collecting Carry out pretreatment.Face datection and the requisite link of face recognition process are the pretreatment of facial image, face figure As the quality of quality is directly connected to the accuracy rate of figure identification.Because the environment of image acquisition is extremely complex, cause what is collected Image incorporates noise, so as to cause distortion.In order to ensure the quality of image, the pretreatment of image is necessary.Conventional face The method of Image semantic classification has：Greyscale transformation, binaryzation, the normalization of image, image filtering, image sharpening etc..

Step 3, the feature extraction based on core kernel linear discriminant analysis (KNLDA)

Core kernel linear discriminant analysis are the improvement based on LDA methods and kernel function in feature extraction.Linear discriminant Analysis (LDA) is widely used to the fields such as machine vision and image recognition, but is difficult to the small sample problem for solving to run into, and zero is empty Between linear discriminant analysis (NLDA) diagnostic characteristicses of sample are extracted in the kernel of the total within class scatter matrix of training sample, gram The small sample problem for having taken, improves discrimination, however, NLDA remains a kind of extracting method of linear character, it is impossible to effectively Extract the nonlinear characteristic of sample.By nonlinear mapping, it is empty that input space sample is mapped to high dimensional feature by Kernel-Based Methods Between, Linear feature extraction algorithm is utilized in high-dimensional feature space, so as to effectively extract the nonlinear characteristic of sample.In practical stability Face identification system in, intensity of illumination, human face posture change, shelter problem are a difficult problems for a great challenge all the time, these The error that problem causes may interfere with the feature extraction of some positions of image, while also can to the amplitude of some useful informations Produce impact.The noise jamming that these interference factors bring to the feature extraction of other unscreened parts while also can bring not The impact of profit.So the present invention adopts core kernel linear discriminant analysis feature extracting method.

Using core kernel linear discriminant analysis feature extracting method, its step is as follows：

Assume that C class samples are respectively A¹, A¹, A¹..., A^c, it is n per class sample number_i(i=1,2,3...C), sum is N；Represent i-th (i=1,2,3 ..., C) class, jth (j=1,2,3 ..., n_i) individual sample, then in the class of the i-th class sample Average isC class grand mean of sample is

Linear discriminant analysis (LDA)

Linear discriminant analysis be to extract the low-dimensional feature of higher-dimension sample most distinguishing ability for the purpose of.By in the total class of sample Scatter matrix Sw and total inter _ class relationship matrix Sb are respectively defined as：

Then sample population variance degree matrix St is：

Assume W_LDAFor the projection matrix of LDA algorithm, then LDA rule definitions are:

As projection matrix W_LDAWhen meeting the criterion function maximum of LDA, the linear discriminant feature after sample projection has optimal Resolution capability.W_LDACan be by solvingCorresponding characteristic vector w of the larger eigenvalues of p₁,w₂...,w_d,...w_pComposition is thrown Shadow matrix, however, S_wUsually singular matrix, its order much smaller than sample dimension m so that W_LDACannot solve, that is, occur little Sample problem.

In order to solve small sample problem, using NLDA algorithms, the kernel of St is removed first, by order from all samples to St Space projection, reduces sample dimension without losing the kernel authentication information of Sw, the order of St and is usually N-1, therefore, sample Meet l to the dimension l after the rank space projection of St<N-1<<M, so as to solve the problems, such as the dimension disaster of Sw kernels.Specifically retouch State as follows：

1) kernel of St is removed, it is assumed that U is the projection matrix of the corresponding characteristic vector composition of St nonzero eigenvalues, then have S_w'=U^TS_wU and S_b'=U^TS_bU。

2) S is calculated_w' kernel, it is assumed that P is S_w' kernel, then have S_w"=P^TS_w' P=(UP)^TS_w(UP)=0 He S_b"=P^TS_b' P=(UP)^TS_b(UP).Wherein UP is effective kernel of Sw, and Extraction and discrimination feature is contributed.

If 3) S_b" kernel exist, then remove it, i.e., in S_w' kernel in maximize sample between class distance, Assume that V represents S_b" nonzero eigenvalue corresponding characteristic vector composition matrix, then the projection matrix of NLDA be represented by W_NLDA= UPV。

1) and 2) through, S_b" non-singular matrix is usually, therefore the 3rd step can be omitted, now NLDA projection matrixes can be represented For W_NLDA=UP, wherein W_NLDARepresent the projection matrix of NLDA.

NLDA algorithms still can not effectively extract the nonlinear characteristic of sample, and kernel method can pass through nonlinear mapping, Input space sample is mapped to into high-dimensional feature space, the non-thread of sample most distinguishing ability is then extracted in high-dimensional feature space Property feature.Present invention introduces the advantage of kernel method, introduces KNLDA algorithms, and recognition of face is applied to, will high dimensional feature sky Between the total within class scatter matrix of middle sample kernel projection matrix as recognition of face projection matrix.Through nonlinear mapping Φ, by input space sampleIn transforming to high-dimensional feature space F, per average of the class sample after nonlinear mapping ForAssume W_KNLDA=[w₁,w₂,...,w_d,...,w_p] for the projection matrix of KNLDA, then it is special in higher-dimension In levying space F, sampleFeatureIt is represented by：

In high latitude feature space F, sample characteristicsDiscrete matrix in total classIt is expressed as：

WhereinRepresent the i-th class sample characteristics in high-dimensional feature space FAverage, WhereinFor total within class scatter matrix in sample high-dimensional feature space F, However, the concrete form of nonlinear mapping Φ is difficult to obtain, it is thus impossible to directly by rightCarry out Eigenvalues Decomposition to try to achieve The projection matrix W of KNLDA_KNLDA。

According to kernel function theory, in high-dimensional feature space F, projection matrix W_KNLDAAphylactic map projection vector Wd is deployable is All training sample sums after nonlinear mapping Φ, i.e.,

Wherein,Expression projection vector Wd refers to training towards the expansion coefficient after projecting direction projection, here projecting direction SampleThe direction formed Jing after nonlinear mapping ΦDefine kernel functionWill Sample in high-dimensional feature space FWith every class sample averageProject to Wd respectively, obtain

HaveWherein a=[a₁,a₂,...,a_d,...,a_p] for projection vector towards The coefficient matrix of expansion coefficient composition after projecting direction, here, projecting direction refers to what training sample was formed Jing after nonlinear mapping Direction,D=1,2,3, ..., P,Therefore can WillAbbreviation is：

Wherein,Dimension size is N × N, and a can be obtained by solving the eigenvalue and characteristic vector of Q, by The corresponding characteristic vector of the p of the Q of selection less eigenvalue is constituted, and dimension size is N × p, while p ＜ ＜ N, therefore, In feature space F, sampleFeature be represented by again：

WhereinIt is the sample after nonlinear mapping ΦTo W_KNLDAThrowing Shadow vector Wd projects extracted feature, is described in detail below：

1) input space sample is mapped in high-dimensional feature space F, i.e.,

2) design factor matrix a, first, calculating matrixThen Eigenvalues Decomposition is carried out to Q, is selected Take the characteristic vector composition matrix a corresponding to the p less eigenvalue of Q

3) sample matrix, sample are calculatedFeature beWherein

From the foregoing, it will be observed that KNLDA algorithms are by introducing kernel functionObtain low-dimensional square Battle array, chooses the P less eigenvalue character pair vector composition a of Q, and the dimension size of matrix a and Q is respectively N × p and N × N, Wherein p ＜ ＜ N, therefore, kernel method adopts Mercer kernel functions by inner product form, effectively prevent and directly calculate the non-of complexity Linear mapping function Φ, solves higher-dimension within class scatter matrixDimension disaster problem.

Step 4, Training RBF Neural Network, the basic thought of RBF neural is：With RBF as implicit unit " base " Constitute implicit sheaf space, the subspace that the spatial decomposition of input is formed into many hyperspheres.In order to carry out to these subspaces Classification generally needs to use clustering algorithm, and conventional has hierarchical clustering, Fuzzy C-Means Clustering and C- mean clusters etc. without prison The learning algorithm superintended and directed, the common feature of these algorithms is that the classification information of training sample is not used in the middle of the classification of sample. RBF neural employs the learning algorithm of supervision.Then the subspace of decomposition is mapped directly to into implicit sheaf space, so as to Efficiently solve the problems, such as in lower dimensional space linearly inseparable, by the way that lower dimensional space is converted to into higher dimensional space so that phenomenon can not The problem divided becomes linear separability.RBF neural is more accurate than the metric form of Euclidean distance.

The training study of RBF networks is generally divided into two processes：

First process is to determine center and the variance of hidden layer, and the selection at center is to mapping space between input layer and hidden layer Structure have a significant impact；

Second process is determination of the hidden layer to connection weight between output layer.

Because the output of output layer is obtained by the output linearity weighted array of hidden node, therefore the output of network is only by few The impact of number weights, needs the weights for adjusting also relatively fewer in training process, greatly reduce the learning time of RBF networks.

The method of RBF function Selection Centers has many kinds, RBF Web vector graphic Orthogonal Least Squares (OLS) of the present invention Learning algorithm.It is chosen to reducing the larger vector of error contribution as in network by orthogonalization method in input vector The heart.

If the training sample set of network is { X₁,X₂,...,X_N, the desired output of network is { d₁,d₂,...,d_N, N is represented The number of training sample, RBF networks may be used to lower linear equation and represent：

M is the number of hidden node in formula, can represent RBF networks with the matrix form of formula：

D=P Θ+E

If the least square solution of above formula isTo regression matrix P Orthogonal Decomposition P=WA, in formula, A is upper three angular moment of M × M Battle array, the value of the element on diagonal is that 1, W is the orthogonal matrix of N × M, there is H=W^TW, H are diagonal matrix, another A θ=g then d= Wg+E.The least square solution of g is expressed as in formula：

Wherein,Orthogonal Decomposition is carried out to P using Schmidt process, in the process for selecting character subset In, base vector Wi is determined using the maximum criterion of error ratio, so that it is determined that the center of network, error ratio is defined as follows：

According to range of error ρ set in advance so as to meet following conditionNever hidden node is determined Number M.The selection of RBF network centers number has a great impact to the performance of network.Orthogonal Least Square can be selected automatically Select network center, it is to avoid randomly select the problem of center band.

Because recognition of face belongs to pattern classification, the knowledge of face is realized using the generalized network GN in RBF network modeies Not.RBF network input layers node number is equal to the dimension of face feature vector, exports the classification of the number for sample of node layer Number.The selection of Basis Function Center and the node number of hidden layer are automatically determined by Orthogonal Least Square.Hidden node parameter once It is determined that, so that it may obtain exporting weights by solving system of linear equations.The knowledge to face can be realized using integrated RBF networks Not.

Step 5, system are realized and recognition result

The identity for identifying people to be identified is the last stage of recognition of face, and each sub-network is only responsible for one class of identification Pattern, the output node number of each sub-network is 1, needs to train K (K is the classification number of sample) individual sub-networks to constitute altogether Integrated network.Exporting come the generic of discriminating test sample according to all-ones subnet network, is finally completed identifying purpose.The present invention With reference to OpenCv computer visions storehouse and the softwares of Visual Studio 2010, a real-time recognition of face system is have developed System.The main interface of the identifying system is as shown in Figure 4.Face identification system designed by the present invention is mainly examined automatically including face Survey, face information is input into, face identity differentiates three modules.When program starts to perform, system will automatically load Face datection Module, photographic head is opened and face is detected in specific region.Click on " Enrol " to be registered, there is provided herein three Logon mode is planted, is respectively：Shoot, register from the registration of existing picture library, from template file.We are shot using photographic head and obtain head Picture, as shown in figure 5, click on " Train " being trained, its objective is to set up training image collection, be that recognition of face is prepared.Fig. 6 For example training sample set.If after the completion of face database training, next just the face in video can be identified, Identity discriminating can be carried out to face to be detected.If face to be identified is judged as the face in face database, in inspection Survey the name information that frame upper left side shows face to be detected.Face identification system run original state design sketch as shown in fig. 7, Training surface chart is as shown in Figure 8.The present invention design face identification system expression shape change it is larger, exist block interference under carry out Experiment, its objective is to pass judgment on the face recognition algorithms after improvement.As a result show, the present invention changed greatly in human face expression, Presence block wait disturb when, still keep higher discrimination, as shown in Figure 9.

Claims (4)

1. a kind of RBF neural face identification method based on KNLDA, its step is as follows：

1) facial image is read, is set up under YCbCr color spaces based on the statistics mould of the colour of skin using human body complexion Clustering features Type, from altimetric image to be checked human face region is filtered out；

2) pretreatment is carried out to human face region, best projection matrix is obtained using the KNLDA feature extracting methods based on kernel function；

3) training of RBF neural is carried out on the sample of feature extraction；

4) specimen discerning is carried out with RBF neural.

2. a kind of RBF neural face identification method based on KNLDA as claimed in claim 1, it is characterised in that step The NLDA feature extracting methods based on kernel function described in 2, comprise the following steps that：

1) input space sample is mapped in high-dimensional feature space F, i.e.,

2) design factor matrix a, first, calculating matrixThen Eigenvalues Decomposition is carried out to Q, chooses Q's P less eigenvalues corresponding to characteristic vector composition matrix a；

3) sample matrix, sample are calculatedFeature beWherein

3. a kind of RBF neural face identification method based on KNLDA as claimed in claim 1, it is characterised in that step RBF is trained using Orthogonal Least Square (OLS) learning algorithm in 3, it chooses right by orthogonalization method in input vector The larger vector of error contribution is reduced as network center；

If the training sample set of network is { X₁,X₂,...,X_N, the desired output of network is { d₁,d₂,...,d_N, N represents training The number of sample, RBF networks may be used to lower linear equation and represent：

$d\left(n\right)=\underset{i=1}{\overset{M}{\&Sigma;}}{p}_i\left(n\right){\mathrm{\&theta;}}_i+\mathrm{\&epsiv;}\left(n\right),n=1,2,...,N$

M is the number of hidden node in formula, can represent RBF networks with the matrix form of formula：

D=P Θ+E.

4. a kind of RBF neural face identification method based on KNLDA as claimed in claim 1, it is characterised in that step RBF neural grader is adopted in 3, is identified and the immediate facial image of images to be recognized, reach the mesh of identity discriminating 's.