CN106446774A - Face recognition method based on secondary nearest neighbor sparse reconstruction - Google Patents

Abstract

The invention discloses a face recognition method based on secondary nearest neighbor sparse reconstruction. The face recognition method comprises the following steps of: firstly, in an original training sample, sample collaborative representation which is more similar to a target sample is selected so as to shorten operation time to a certain degree; and after a coding coefficient is obtained, classification is not directly carried out, instead, the target sample is subjected to primary reconstruction to obtain a reconstruction sample which has the same category number with the original sample, and the sample collaborative representation which is similar to the target sample is selected from the reconstruction sample again. Operation efficiency can be further improved, a classification target is further reduced through the secondary reconstruction of the target sample, and recognition is more accurate.

Description

A kind of face identification method based on the reconstruct of secondary neighbor sparse

Technical field

The present invention relates to a kind of face identification method, particularly a kind of recognition of face side based on the reconstruct of secondary neighbor sparse Method.

Background technology

Nowadays, in traffic and transport field, particularly higher airport, station etc. is required all to pass through computer safe class Visual system carries out real-time security monitoring, and any extraction to correlated characteristic in video and analysis are all based on face.Therefore, robust Property recognition of face is study hotspot and the difficult point in current living things feature recognition field.The method that feature based extracts among these, than As principal component analysiss (PCA), two-dimensional principal component analysis (2DPCA) and linear discriminant analysis (LDA), it is intended to find target image Low-dimensional feature with classification dependency.On this basis, for light, angle, the increasingly complex environment such as even block, very Multi-expert scholar proposes new technology and theoretical, the also true performance improving recognition of face to a certain extent.But so far also Do not have the transformation criterion to lower dimensional space for the authoritative dimensional images.Wherein it is based on PCA and extract feature, due to must be by two dimensional image square Battle array is converted to a dimensional vector, destroys the structural information of image；Though 2DPCA is directly based upon two dimensional image matrix and extracts spy Levy, but the local part critically important to recognition of face can not be extracted；And LAD usually faces " small sample " problem, in single training sample In the case of this, this algorithm just fails.In recent years, the rarefaction representation based on compressed sensing coding theory (SRC) is because making an uproar to image Sound is insensitive and causes extensive concern.This model is proposed by WRIGHT etc. at first, and it passes through in higher dimensional space to facial image Estimate to complete pattern classification.Additionally, some scholars pass through embedded iteration weight coefficient in sparse solution it is proposed that robustness Higher method.Because SRC is with l₁- norm solves sparse coefficient, needs iteration, causes the complexity of calculating higher.To this Yang and Zhang proposes to extract image local direction characteristic for SRC based on Gabor transformation, reduces the complexity of algorithm, And recognition effect is more preferably.

Generally in actual applications, the sample for training is often incomplete, is so solving coefficient using SRC When, even if target sample determines ownership, limited genus class sample also is difficult to linear correlation.Therefore, ZHANG et al. is referred to by analysis Go out the effect for rarefaction representation for the similarity of sample between class, then propose collaborative presentation class (CRC), this algorithm is based on l₂- model The sparse solution of number, while significantly shortening run time, still keeps and the suitable recognition effect of SRC.On this basis, There are a lot of improvement projects.Utilize the low-dimensional feature of image to implement CRC classification than if any scholar, improve discrimination.LIN proposes Shandong Rod is collaborative to represent (RCR), and the method computation complexity substantially reduces.Additionally, TIMOFT, WAQAS, LU point out effective office of sample The importance to rarefaction representation for portion's information, builds weighting matrix by the similarity of training sample and target sample, is embedded in and is During number solves, propose weighting rarefaction representation respectively, effectively enhance classification performance.But the method is based on whole data set and works in coordination with table The similarity of each training sample and target sample before showing, to be weighed, the efficiency of operation can be reduced to a certain extent.The pins such as ZHANG The information such as the illumination in sample, angle, attitude are unable to the interference to classification for the effectively utilizes, construct virtual sample by different scenes Originally it is embedded in all kinds of training samples although there being preferable recognition effect, but the training sample of expansion undoubtedly can consumption coefficient be asked The time of solution.XU etc. reduces target class other secondary classification method by stage makes identification more accurate, but is equally based on and entirely counts Collaborative expression upper not too big advantage at runtime according to collection.

To sum up the analysis based on CRC method popular in recent years, all has common trait, that is, be all based on whole data set Carry out collaborative expression, pattern classification is completed by reconstruct.Here there are two problems：(1) face database is larger, uses all samples The collaborative speed representing, coefficient solution will certainly being reduced；(2) similar image inevitable exist block, if both using there being screening Gear also removes collaborative expression testing image using unscreened sample, then would necessarily affect the effect of expression.And ZHANG etc. tests Demonstrate,prove only suitable sample size and could obtain higher discrimination.

Content of the invention

The technical problem to be solved is to provide a kind of insensitive for noise that image itself is existed and overall knowledge The all higher secondary neighbor sparse reconstruct recognition of face new method of other performance.

For solving above-mentioned technical problem, the technical solution adopted in the present invention is：One kind is based on secondary neighbor sparse reconstruct Face identification method it is characterised in that comprising the steps of：

Step1：Face database contains the image of C people, and everyone has n_iWidth image, each image size be m × N, defines training sample setTest sample is Y ∈ R^m×n, will be all kinds of Training sampleVector turns toSuch i-th class training sample composition matrix beThe matrix that C class training sample is constituted is Χ=[Χ₁,…,X_C]∈R^m×N, by test sample Y Vector turns to y ∈ R^m×1；

Step2：Calculate each training sampleThe distance between with y, and find out in all kinds of training samples with test sample y away from From K nearest neighbour's sample, the matrix of i-th class K neighbour's sample composition isWherein p₁, p₂,…,p_K∈[1,n_i], then new neighbour's sample matrix

Step3：Represent model based on collaborative, willAs encoder dictionary, solve code coefficient α with method of least square, obtain

Step4：It is associated with all kinds of training samples in usage factor vector αSparse coefficient α_iReconstruct test sample respectively Y, obtains belonging to all kinds of reconstructed samplesThen build new reconstructed sample collection, that is,

Step5：Based on new reconstructed sample spaceCalculate all kinds of reconstructed samplesWith the distance of test sample y, and look for Go out K' wherein closest with y reconstructed sampleThen new neighbour's reconstructed sample matrix Wherein q₁, q₂..., q_K'∈[1,C]；

Step6：It is again based on collaborative expression model, willAs new encoder dictionary, solve coding with method of least square Coefficient

Step7：Usage factor vectorIn be associated with each K' neighbour's reconstructed sampleSparse coefficientWeigh again respectively Structure test sample y, obtains secondary reconstructed sampleWherein q_K'∈[1,C]；

Step8：Calculate secondary reconstructed sample respectivelyError with test sample yI.e.

Step9：According to reconstructed errorMinima judge the ownership of test sample y.

Further, the original training sample with test sample y neighbourAnd a reconstructed sampleAll by European away from From i.e. formulaWithIt is calculated；

Further, in first time collaborative expression, not directly according to test sample y and a reconstructed sampleMistake Difference class, but select the K' reconstructed sample similar to test sample y againFor secondary reconfiguration classification.Wherein q_K' ∈[1,C]；

Further, during secondary collaborative presentation class, test sample y and K' neighbour's reconstructed sampleErrorPass through Euclidean distanceIt is calculated, rather than traditional calculations in synergetic classificationMode, that is,

Further, the described ownership judging test sample y then passes through formula Differentiate.

The present invention compared with prior art, has advantages below and effect：

1 in view of there are a lot of local similar information in each training sample and test sample, and such synergetic classification is to a certain degree On influence whether classify actual effect.Rarefaction representation unstable the asking of classification can be solved by selecting more effective sample to work in coordination with expression Topic；

2nd, select and the more like training sample of test sample is worked in coordination with and represented, improve to a certain extent based on l₂- model Number solve coefficients speed, and this by the sparse reconstruct of neighbour's sample classification when still have preferable recognition effect；

3rd, the sample second selecting neighbour based on sparse reconstruct for the first time, reduces classification range so that identification is more accurate, And further shorten run time.

Specific embodiment

Below by embodiment, the present invention is described in further detail, following examples be explanation of the invention and The invention is not limited in following examples.

The present invention passes through to select and the more like instruction of target sample based on the face identification method that secondary neighbor sparse reconstructs Practice sample work in coordination with represent, can reduce coefficient solve the time, target sample is once reconstructed, from original sample classification number In identical reconstructed sample, second selecting and the close sample of target sample are worked in coordination with and are represented, can reduce class object, make identification More accurate, and further increase operation efficiency.Method comprises the following steps that：

Step1：If face database contains the image of C people, everyone has n_i(i=1,2 ..., C) width image, by This defines training sample set It is the i-th class jth width image pattern, Each image size is m × n, and test sample is Y ∈ R^m×n, by all kinds of training samplesVector turns toSuch I class training sample composition matrix beC class training sample constitute matrix be Χ= [Χ₁,…,X_C]∈R^m×N, test sample Y vector is turned to y ∈ R^m×1.

Step2：Calculate the i-th class training sample with Euclidean distanceDistance with test sample yI.e.Wherein i=1,2 ... C, j=1,2 ..., n_i.

Step3：The K neighbour sample closest with test sample y is found out in the i-th class training sampleComposition the I class neighbour's sample matrixWhereinI=1,2 ..., C, j=1,2 ..., n_i, p₁,p₂,…,p_K∈[1,n_i].The matrix that then new neighbour's sample is constituted is

Step4：WillAs the encoder dictionary of collaborative presentation class, solve code coefficient α with method of least square, obtainWherein λ is regularization parameter, to keep the stable and openness of sample reconstruct.

Step5：It is associated with all kinds of neighbour's samples in usage factor vector αSparse coefficient α_iEach via Self-reconfiguration test sample Y, must belong to all kinds of reconstructed samplesI.e.Wherein i=1,2 ... C, in formulaFor i-th class neighbour's sample Reconstruct.Then C reconstructed sampleComposition matrix be

Step6：Calculate the i-th class reconstructed sample with Euclidean distanceDistance with test sample yI.e.Wherein i=1,2 ... C.

Step7：In C reconstructed sampleIn find out the K' neighbour sample closest with test sample yWhereinWherein i=1,2 ..., C, q_K'∈ [1, C], then K' neighbour's reconstructed sampleComposition matrix beWherein q₁, q₂..., q_K'∈[1,C].

Step8：It is again based on collaborative expression model, willAs encoder dictionary, solve code coefficient with method of least square?

Step9：Usage factor vectorIn be associated with K' neighbour's reconstructed sampleSparse coefficientSecondary heavy respectively Structure test sample y, obtains each neighbour's reconstructed sampleI.e.Wherein q_K'∈ [1, C], in formulaFor Secondary neighbour's sampleReconstruct.

Step10：Test sample y and each neighbour's reconstructed sample are calculated according to Euclidean distanceErrorI.e.Wherein q_K'∈[1,C].

Step11：Differentiate the ownership of test sample y according to following formula, that is,

There is certain defect in the rarefaction representation based on whole data set, exist in actual applications for lifting CRC further Performance in recognition of face, designs a kind of sparse reconfiguration classification method based on secondary neighbour.By simulation results show, the present invention Face identification method have greatly improved on recognition performance.The method finds test sample first on original sample collection Neighbour, only selects to be suitable for quantity and effective sample is worked in coordination with and represented, improve the speed of coefficient solution to a certain extent；So Afterwards respectively with the sparse reconstruct sample to be tested of all kinds of neighbour's samples, obtain and original sample classification number identical reconstructed sample；Then Find the neighbour of test sample based on reconstructed sample collection again, and collaborative expression, carry out secondary reconstruct, final implementation pattern divides Class.This by secondary neighbour select the sparse reconstructing method of more effective sample reduce class scope further, classification can be made More accurate, shorten run time simultaneously.

Above content described in this specification is only illustration made for the present invention.The affiliated technology of the present invention The technical staff in field can be made various modifications or supplement or substituted using similar mode, only to described specific embodiment Without departing from the content of description of the invention or to surmount scope defined in the claims, the guarantor of the present invention all should be belonged to Shield scope.

Claims (5)

1. a kind of face identification method based on the reconstruct of secondary neighbor sparse is it is characterised in that comprise the steps of：

Step1：Face database contains the image of C people, and everyone has n_iWidth image, each image size is m × n, definition Training sample setTest sample is Y ∈ R^m×n, by all kinds of training samples ThisVector turns toSuch i-th class training sample composition matrix beC class is instructed Practice sample constitute matrix be

Χ=[Χ₁,…,X_C]∈R^m×N, test sample Y vector is turned to y ∈ R^m×1；

Step2：Calculate each training sampleThe distance between with y, and find out in all kinds of training samples with test sample y distance K near neighbour's sample, wherein i-th class k nearest neighbor sample composition matrix be

Then new neighbour's sample matrix

Step3：WillAs encoder dictionary, solved based on the collaborative code coefficient α representing model with method of least square, obtain

Step4：It is associated with all kinds of neighbour's samples in usage factor vector αSparse coefficient α_iReconstruct test sample y respectively, obtain To belonging to all kinds of reconstructed samplesThen build new reconstructed sample collection, that is,

$\stackrel{\‾}{Y}=\{{\stackrel{\‾}{y}}_1,{\stackrel{\‾}{y}}_2,\mathrm{...},{\stackrel{\‾}{y}}_i,\mathrm{...},{\stackrel{\‾}{y}}_C\};$

Step5：Based on new reconstructed sample spaceCalculate all kinds of reconstructed samplesWith the distance of test sample y, and find out it In the K' reconstructed sample closest with yThen new neighbour's reconstructed sample matrix $\tilde{Y}=\[{\tilde{y}}_{q_1},{\tilde{y}}_{q_2},\mathrm{...},{\tilde{y}}_{q_{K^{\′}}}\];$

Step6：WillAs new encoder dictionary, solved based on the collaborative code coefficient representing model with method of least square?

Step7：Usage factor vectorIn be associated with K' neighbour's reconstructed sampleSparse coefficientReconstruct test again respectively Sample y, obtains secondary reconstructed sample

Step8：Calculate secondary reconstructed sample respectivelyError with test sample yI.e.

Step9：According toJudge the ownership of test sample y.

2. according to described in claim 1 based on secondary neighbor sparse reconstruct face identification method it is characterised in that：With test The original training sample of sample y neighbourAnd a reconstructed sampleIt is formula all by Euclidean distanceWithIt is calculated.

3. according to described in claim 1 based on secondary neighbor sparse reconstruct face identification method it is characterised in that：First In secondary collaborative expression, not directly according to test sample y and a reconstructed sampleError classification, that is, skipBut select the K' reconstructed sample similar to test sample y againFor two Secondary reconfiguration classification.

4. according to described in claim 1 based on secondary neighbor sparse reconstruct face identification method it is characterised in that：In reconstruct During classification, test sample y and secondary neighbour's reconstructed sampleErrorBy Euclidean distance, that is,Meter Obtain, rather than traditional calculations in synergetic classificationMode, that is,

This allows for training sample in the case of randomly selecting, and is calculated using Euclidean distance Reconstructed error can obtain more preferable recognition effect.

5. according to described in claim 1 based on secondary neighbor sparse reconstruct face identification method it is characterised in that：Described sentence The ownership of disconnected test sample y passes through formulaJudge, reconstructed by secondary neighbour The corresponding q of minimum error_K'Value, obtains the corresponding i value when once reconstructing, that is, affiliated classification.