CN102324031A - Latent semantic feature extraction method in aged user multi-biometric identity authentication - Google Patents

Abstract

The invention relates to a latent semantic feature extraction method in aged user multi-biometric identity authentication. Identity authentication is performed by performing multi-mode latent semantic analysis and data mining mapping on aged user multi-biometric images and extracting the latent semantic features of the images. According to the latent semantic feature extraction method in the aged user multi-biometric identity authentication, multiple local bottom features can be acquired by extracting the multi-biometric images of face, multiple fingerprints and palm prints and the like; the extracted features can be processed by using a multi-mode latent semantic analysis algorithm on three aspects of bottom feature-image matrix construction, two-dimensional matrix decomposition and clustering algorithm; the processed features are further mined and mapped through an 'intelligent black box model', so that the latent semantic features of the images can be effectively acquired; and the system is automatically adjusted by introducing an adaptive feedback structure with a genetic algorithm (GA), so that modification of the latent semantic features of the images is realized.

Description

Implicit expression semantic feature method for distilling in the authentication of aged user's multi-biological characteristic

Technical field

The present invention relates to technical field of biometric identification, the implicit expression semantic feature method for distilling in the authentication of especially aged user's multi-biological characteristic.

Background technology

At present; China more than 60 years old population reach 1.8 hundred million people, account for total population 13.8%, weigh by international standard; China has got into the society of the aged; Along with country accelerates to set up and improve the social security system that covers urban and rural residents energetically, like the golden granting of social old-age insurance, supplementary pension, medical insurance etc., aged user will become the main colony of Future Society public service; Exist deception, false claiming phenomenon to become social now question of common concern in the distribution process such as social old-age insurance gold, supplementary pension, informationization, digitizing, network technology provide help for solving aged authenticating user identification quagmire.At present, biometrics identification technology, long-distance video authentication are falsely claimed as one's own aged user's in the phenomenon identity by successful Application to examining the social pension gold.

Biometrics identification technology carries out identification and checking through utilizing intrinsic physiological characteristic of human body and behavior act.According to the kind and the number that use biological characteristic; Living things feature recognition can be divided into single living things feature recognition and multi-biological characteristic identification; As using the widest single bio-identification identity identifying technology, the problem of authentication had received extensive concern when fingerprint recognition was put at the aged user's social pension golden hair of solution.As far back as 1901, Britain began employing fingerprint and has discerned and avoid the railway worker to falsely claim as one's own, lead more salary.At present, associated companies such as IBM, Microsoft, HP, Compaq, Changchun letter reach, are just waiting in the Hangzhou service field that entered society of the product of company." the payment pension fingerprint identity validation systems technology standard (trying) " of the social insurance career management center issue of China Ministry of Labour and Social Security also will issue as the social public service standard based on the fingerprint identification method of minutiae point (minutiae); But; Concerning aged user; Owing to have experienced all sorts of hardships, fuzzy finger is very common, and traditional fingerprint recognition system based on minutiae point tends to cause system's misclassification rate to increase because the extraction minutiae point is undesirable even authentication was lost efficacy.In addition; Recognition technology based on single biological characteristic exists not ubiquity: some biological characteristic disappearance (like severed digit), damage (like impaired finger), pathology (like cataract) or collection apparatus second-rate (changing like people's face light) all can cause robustness, the poor reliability of recognition system; A little less than the anti-duplicity, be difficult to satisfy the actual requirement of different occasions.

Image implicit expression semantic feature (Image Latent Semantic Features; ILSF) obtain by low-level image feature-image array; Have the information abundanter more than traditional image, semantic, but relative and low-level image feature, and these characteristics have stronger expression and classification capacity.Therefore, the characteristic of utilizing TLSA to extract can be used as the characteristic of a kind of " uniqueness ", and is proved to be and can be used in the biometric identity field of authentication.Simultaneously, compare traditional low-level image feature, owing to be used for describing image indirectly; Image implicit expression semantic feature is not very high for the quality requirements of images acquired; Can better overcome the influence that some unfavorable factor is brought, fuzzy such as the image streakline of fingerprint, and the influence of human face light variation.

Summary of the invention

The technical matters that the present invention will solve is: in order to overcome the above-mentioned middle problem that exists; Implicit expression semantic feature in the authentication of a kind of aged user's multi-biological characteristic method for distilling is provided, utilizes image processing techniques and intellectual technology the user to be carried out the technology of authentication.

The technical solution adopted for the present invention to solve the technical problems is: the implicit expression semantic feature method for distilling in the authentication of a kind of aged user's multi-biological characteristic; Shine upon through aged user's multi-biological characteristic image being carried out semantic analysis of multimode implicit expression and data mining, and extract image implicit expression semantic feature and carry out authentication.

Described multi-biological characteristic image comprises people's face, refers to fingerprint and palmmprint more that the concrete steps of the multiple local low-level image feature of its extraction multi-biological characteristic image are following:

A. biometric image pre-service: refer to that by people's face, four emerging system that fingerprint and palmmprint constitute carries out pre-service;

B. extract low-level image feature: extract invariant moment features, Garbor filter characteristic, direction equalization characteristic and half-tone information entropy characteristic;

C. extract the local low-level image feature in the low-level image feature: successively through to the selection of RP, based on the extraction of the ROI of RP and the division of ROI; Through extracting the local low-level image feature of image, the local bottom that extracts image is characterized as invariant moment features, Garbor filter characteristic, direction equalization characteristic and half-tone information entropy characteristic at last.

The concrete steps of described multimode implicit expression semantic analysis are following:

A. the image array of low-level image feature makes up: adopt multiple low-level image feature, make up the image array of each user's low-level image feature;

B. parallel two-dimentional nonnegative matrix decomposition algorithm: the image array to low-level image feature carries out the diagonalization processing earlier; Again diagonalizable matrix being carried out the row matrix direction decomposes; And then former diagonalizable matrix is carried out transpose process obtain column direction information, the basis matrix that obtains is carried out basis matrix orthogonalization;

C. fuzzy C-means clustering: utilize Fuzzy C average (FCM) clustering method in the programming tool case to carry out quick clustering.

The concrete steps of the data after the semantic analysis of multimode implicit expression being carried out the data mining mapping are following:

A. beta pruning Algorithm Analysis: the network structure predefine parameter of initialization fuzzy neural network (FNN) (like convergence constant a, attenuation constant b, least error e, regular importance threshold value f); Import first eigenvectors; Produce article one fuzzy rule; Any input feature value is calculated the distance of itself and first eigenvectors, draw minimum value d _MinThereby, calculate actual output error e _iIf, error e _iGreater than regular importance fault value f, then produce new fuzzy rule, thus adjustment network architecture parameters (like convergence constant a, attenuation constant b, least error e, regular importance threshold value f);

B. extract image implicit expression semantic feature: comprise off-line learning stage and on-line testing stage.

The described off-line learning stage is to be used for training fuzzy neural network (FNN) and to adopt the beta pruning algorithm that the network structure of fuzzy neural network (FNN) is done dynamic adjustment through learning sample.

The described on-line testing stage is to utilize the fuzzy neural network that trains that test sample book is tested, thereby it is semantic to extract image implicit expression.

The concrete steps that extraction image implicit expression semantic feature is revised are following:

A. introduce self-adaptation dynamic feedback structure: in intelligent blackbox model, utilize the self-adaptation feedback arrangement of band GA optimized Algorithm to extract the semantic feature and the state parameter of identification based on fuzzy neural network;

B. image implicit expression semantic feature normalization: the state parameter through picking out intelligent blackbox model and with obtain through sample learning decide normal condition and compare; Draw the input of difference as model; Thereby the image implicit expression semantic feature deviation that must cause because of environmental difference, thereby obtain normalized image implicit expression semantic feature.

The beneficial effect of the implicit expression semantic feature method for distilling in aged user's multi-biological characteristic of the present invention authentication is: through extracting people's face, referring to multi-biological characteristic images such as fingerprint and palmmprint more, can obtain multiple local low-level image feature; Utilize multimode implicit expression semantic analysis algorithm to decompose and clustering algorithm three aspects, can handle the characteristic of extracting from low-level image feature-image array structure, two-dimensional matrix; Further the characteristic after handling is excavated mapping through " intelligent blackbox model ", can obtain image implicit expression semantic feature effectively; Through introducing the self-adaptation feedback arrangement of band genetic algorithm (GA), system is adjusted automatically, realize the correction of image implicit expression semantic feature, the good and good reliability of the method acquisition quality can satisfy the actual requirement of different occasions.

Description of drawings

Below in conjunction with accompanying drawing and embodiment the present invention is further specified.

Fig. 1 is a structural principle block diagram of the present invention;

Fig. 2 is the synoptic diagram of two-dimensional matrix diagonalization (a) row combination (b) row combination among Fig. 1;

Fig. 3 is a FCM cluster synoptic diagram among Fig. 1;

Fig. 4 is a self-adaptation dynamic feedback structured flowchart among Fig. 1;

Fig. 5 is the normalized structured flowchart of image implicit expression semantic feature among Fig. 1.

Embodiment

Combine accompanying drawing that the present invention is done further detailed explanation now.These accompanying drawings are the synoptic diagram of simplification, basic structure of the present invention only is described in a schematic way, so it only show the formation relevant with the present invention.

Implicit expression semantic feature method for distilling in aged user's multi-biological characteristic authentication as shown in Figure 1; Refer to that by people's face, four emerging system that fingerprint and palmmprint constitute carries out pre-service respectively; From pretreated biometric image, extract the image low-level image feature of invariant moment features (comprising hu invariant moments and zernike invariant moments), Garbor filter characteristic, direction equalization characteristic and half-tone information entropy characteristic; Image low-level image feature after the extraction at first carries out multiple low-level image feature; Make up the image array of each user's low-level image feature, secondly the image array to low-level image feature carries out the diagonalization processing, again diagonalizable matrix is carried out the row matrix direction and decomposes; And then former diagonalizable matrix is carried out transpose process obtain column direction information; Basis matrix to obtaining carries out basis matrix orthogonalization, utilizes the fuzzy C-means clustering method in the programming tool case to carry out cluster again, after the data after the semantic analysis of process multimode implicit expression are carried out the data mining mapping; In intelligent blackbox model, utilize the self-adaptation feedback arrangement of band GA optimized Algorithm to extract the semantic feature and the state parameter of identification based on fuzzy neural network.

The concrete steps of the implicit expression semantic feature method for distilling in aged user's multi-biological characteristic of the present invention authentication are following:

One, the extraction of low-level image feature: comprise the biometric image pre-service, extract low-level image feature and extract three key steps of local low-level image feature:

(1) biometric image pre-service: the pre-service of multi-biological characteristic image is one of committed step before the feature extraction; Because emerging system refers to that by people's face, four fingerprint and palmmprint etc. constitute; Therefore need carry out pre-service respectively to it, pretreated key step comprises: area-of-interest (ROI) is cut apart, enhancing, normalization etc.At first, we will extract the ROI of image: 1. to people's face, mainly from video, detect and cut apart facial image; 2. to the hand images acquired; To cut apart earlier and refer to fingerprint and palmmprint more; Through four fingers and palmmprint are located, the thinking (document sees reference) of reference Uhl etc. refers to that with four fingerprints, palmmprint split one by one; Then, on the basis of early-stage Study, the ROI of images such as people's face of obtaining, fingerprint, palmmprint is strengthened respectively and normalization is handled;

(2) extraction of low-level image feature: get invariant moment features (hu invariant moments and zernike orthogonally-persistent square), Garbor filter characteristic, direction equalization characteristic, half-tone information entropy characteristic etc. indescribably.

(a) invariant moment features: invariant moment features has rotation, yardstick and translation invariant feature, has the provincial characteristics ability of very strong description image.Invariant moments relatively commonly used is like hu invariant moments and zernike orthogonally-persistent square, and its key step is following:

Step 1. is according to the eigenwert I of 7 hu invariant moments of hu invariant moments formulas Extraction;

Step 2. is according to formula $A_{\mathrm{Nm}}=\frac{n+1}{\mathrm{\π}}\underset{x=0}{\overset{M-1}{\mathrm{\Σ}}}\underset{y=0}{\overset{N-1}{\mathrm{\Σ}}}f(x,y)V_{\mathrm{Nm}}^{*}(x,y),$ Select the exponent number n and the repeat number m of zernike invariant moments, extract A _NmValue;

The I and the A of step 3. pair extraction _NmBe combined into invariant moment features.

(b) Garbor filter characteristic: Gabor filter has good direction and selects and frequency selective characteristic, can carry out time frequency analysis to image, extracts the texture value under different directions, the frequency, and its key step is following:

Step 1: according to formula $G(x,y;f,\mathrm{\θ})=\mathrm{Exp}\{-\frac{1}{2}[\frac{x^{\′}}{{\mathrm{\δ}}_{x^{\′}}^2}+\frac{y^{\′}}{{\mathrm{\δ}}_{y^{\′}}^2}\left]\right\}\mathrm{Cos}\left(2\mathrm{\π}{\mathrm{Fx}}^{\′}\right),$ X '=xsin θ+ycos θ,

Y '=xcos θ-ysin θ selects rational direction θ and frequency f parameter (to see that the applicant publishes thesis ^[16]), extract the G value under different directions and the different frequency;

Step 2: the G value to extracting is combined into Gabor filter characteristic.

(c) direction equalization characteristic: the direction equalization has been described among the image direction figure central block and the relation of neighborhood piece on every side; The direction of neighborhood piece around if the direction of central block is more approaching, its value will be high more, therefore; Can be used for describing the direction distribution situation of image, its key step is following:

Step 1: the direction θ that calculates local streakline;

Step 2: according to formula $C(x,y)=\frac{{\mathrm{\Σ}}_{(i,j)\∈W}|\mathrm{Cos}(\mathrm{\θ}(x,y)-\mathrm{\θ}(x_i,y_i))|}{W\×W}$ The calculated direction equilibrium value;

Step 3: the equilibrium value C to extracting is combined into direction equalization characteristic.

(d) half-tone information entropy characteristic: closely related average probabilistic the measuring that can be used as local gray level distribution in the image of information.If certain regional grey scale change is violent or frequent; Explain that this regional picture material is complicated and unstable relatively; Have higher uncertain degree and higher half-tone information moisture in the soil value; On the contrary, lower uncertain degree and half-tone information moisture in the soil value that the zone that picture material is unified and stable is corresponding, its key step is following:

Step 1: according to formula $H=-\underset{i=1}{\overset{M}{\mathrm{\Σ}}}\underset{j=1}{\overset{N}{\mathrm{\Σ}}}{P}_{\mathrm{Ij}}\mathrm{Log}{P}_{\mathrm{Ij}}$ The information entropy of computed image, wherein P _IjBe gradation of image value f (i, the frequency that j) in image, occurs in residing gray scale;

Step 2: the half-tone information entropy H to extracting is combined into half-tone information entropy characteristic.

(3) local bottom feature extraction: extract the concrete information that local feature can accurately be expressed image and comprised, and can overcome the influence of noise and image deformation.Through image being extracted area-of-interest (ROI) and divide, and the extraction of local feature is carried out in the zoning, its key step is following:

Step 1: the selection of RP, the selection of general RP is based on the central point with global characteristics of image, in order to extract ROI and characteristic accurately, needs to confirm unique RP.As fingerprint, refer to the den and slap the central point of root distance based on the distance between two irises, the main foundation of palmmprint according to central point, people's face;

Step 2: based on the extraction of the ROI of RP.Position and direction with RP are that ROI is confirmed at the center, and every width of cloth image all can obtain unique candidate ROI zone and be convenient to Feature Extraction;

The division of step 3:ROI.In order to overcome factors such as non-linear deformation and noise to extracting the influence of characteristic, need be with the ROI piecemeal.Can adopt dual mode: the method that 1. adopts rectangular node extracted region characteristic; 2. adopt annular piecemeal technology to explain image-region, can overcome the shortcoming that the rectangular node field method needs accurate zoning, thereby reduce the time consumption of system, improve discrimination;

Step 4: extract local low-level image feature.Comprise invariant moment features (hu invariant moments and zernike orthogonally-persistent square), Garbor filter characteristic, direction equalization characteristic, half-tone information entropy characteristic etc.

Two, the semantic analysis of multimode implicit expression is handled:

(1) low-level image feature-image array makes up: adopt multiple low-level image feature, make up low-level image feature-image array (q biological characteristic merges altogether) of each user, its concrete steps are following:

Step 1: the ROI image unification of each biological characteristic is blocked into p size be the little image of n * n, q the individual local little image of the common p * q of biometric image;

Step 2: each local little image is comprised analyses such as invariant moment features (hu invariant moments and zernike orthogonally-persistent square), Garbor filter characteristic, direction equalization characteristic, half-tone information entropy characteristic respectively, and with the column vector of these characteristics as low-level image feature-image array of each user;

Step 3: with the row vector of the local little image behind each piecemeal (q biological characteristic) as low-level image feature-image array of each user; Add up the probability of last each low-level image feature of obtaining of step to its appearance; Make up each user's low-level image feature and the characteristic-image array between the image, its size is p * q.

(2) parallel 2D-NNF algorithm: at first; Low-level image feature-image array is carried out diagonalization to be handled; Then, after diagonalizable matrix being carried out the decomposition of row matrix direction, directly former diagonalizable matrix is carried out transpose process and obtain column direction information; At last, the basis matrix that obtains is carried out basis matrix orthogonalization: its key step as shown in Figure 2 is following:

Step 1: diagonalization of matrix: with size is that I is used in m the low-level image feature-image array set of p * q _{P * q}=[S ₁, S ₂..., S _m] represent S _nRepresent low-level image feature-image array of each user, m is a number of users.1), so with the capable combination of matrix, and adopts the mode row combination image of Fig. 2 (a), and obtain diagonalizable matrix A from image array if length p is not more than width q _n(zone shown in the shade).2) if length p greater than width q, is listed as combination with matrix so, and the mode row combination image of employing Fig. 2 (b), and the image array after combination obtains diagonalizable matrix A _n(zone shown in the shade).And the big young pathbreaker who obtains diagonalizable matrix is the same with original matrix, still is p * q;

Step 2: the image array line direction decomposes: with size is that m the diagonalizable matrix of p * q gathered and used X _{P * q}=[A ₁, A ₂..., A _m] represent A _nRepresent the low-level image feature-image array after each user's diagonalization, m is a number of users, and matrix H is long-pending for the matrix L of p * d and size are d * q at first to utilize 1D-NMF to be decomposed into size, makes: X _{P * q}≈ L _{P * d}H _{D * q}Here d is with reference to dimension, and L is that matrix X decomposes the basis matrix that obtains in image row direction, and H is a matrix of coefficients;

Step 3: the image array column direction decomposes: with size is that m the diagonalizable matrix of q * p gathered and used Y _{Q * p}=[B ₁, B ₂..., B _m] represent, wherein

Be that former diagonalizable matrix is carried out transpose process.Similar above-mentioned algorithm utilizes 1D-NMF to find size nonnegative matrix R and nonnegative matrix H that size is r * p for q * r, makes Y _{Q * p}≈ R _{Q * r}H _{R * p}Here, r is with reference to dimension, and R is the basis matrix that the decomposition of matrix Y on image column direction obtains, and H is a matrix of coefficients;

Step 4: matrix X and Y constitute according to the former figure matrix of training sample and transposition figure matrix thereof, therefore can carry out simultaneously their decomposition.And, to diagonalization low-level image feature-image array A of any user _n, its coefficient C on the row and column basis matrix _n=L ^TA _nR, size is d * r, obviously, vectorial dimension reduces greatly.Utilize the basic L of row and Lie Ji R reconstruct low-level image feature-image array to be expressed as: A _n≈ LC _nR ^T, n=1,2 ... m, then two-dimentional basis matrix does

Step 5: basis matrix orthogonalization: the L in the basis matrix

that parallel 2D-NMF method is obtained distinguishes orthogonalization: L '=orth (L) and R '=orth (R) with the R matrix.Two-dimentional basis matrix after the orthogonalization like this Constituted original image matrix A _nAn implicit expression semantic space, a semanteme in the corresponding subspace of every column vector.Project image onto in this semantic space, promptly obtain coefficient C by the semantic feature combination _nRepresented image implicit expression semantic feature.

Three, data mining mapping: based on the design of FNN " intelligent blackbox model "; On the basis of early-stage Study, utilize the FNN construction data to excavate the model of mapping, wherein; Obtaining of sample need be through first three treatment step of Fig. 1; Obtain multimode implicit expression semantic analysis characteristic, and the multimode implicit expression semantic analysis tagsort that obtains is study and trains two groups of samples, at learning phase; Through the beta pruning algorithm network structure of FNN is dynamically adjusted, made up rational FNN " intelligent blackbox model "; At test phase, carry out test sample (like Fig. 3 and shown in Figure 4) with the network parameter that trains:

(1) step of beta pruning algorithm is following:

Step 1: the network structure predefine parameter of initialization FNN (like convergence constant a, attenuation constant b, least error e, regular importance threshold value f), import first eigenvectors, produce the 1st fuzzy rule (if-then mode);

Step 2: any input feature value is calculated itself and the distance of first eigenvectors, searching minimum value d _Min, and calculate actual output error e _iIf, error e _iGreater than regular importance threshold value f; Then produce new fuzzy rule (if-then mode); And calculate the error rate of descent η based on linear regression, otherwise adjustment network architecture parameters (like convergence constant a, attenuation constant b, least error e, regular importance threshold value f) also jumps to step 4;

Step 3:, otherwise return step 2 if rate of descent η, then rejects i bar rule less than least error e

Step 4: return many circulations of step 2, up to all proper vector EOI.

(2) extract image implicit expression semantic feature: FNN is carried out off-line learning and two stages of on-line testing.At learning phase, a large amount of learning samples will be used for training FNN, and adopt the beta pruning algorithm that the network structure of FNN is done dynamic adjustment; At test phase, utilize the FNN that trains that test sample book is tested, extract image implicit expression semantic feature.

Four, image implicit expression semantic feature correction (as shown in Figure 5)

(1) introduces self-adaptation dynamic feedback structure: in " intelligent blackbox model " based on FNN,, utilize the self-adaptation feedback arrangement of band GA optimized Algorithm to extract the semantic feature and the state parameter of identification in the sample learning stage.Its technical scheme is as shown in Figure 4, mainly contains dual mode: 1. FNN optimizes the mode of GA.On the early-stage Study basis, utilize the FL among the FNN dynamically to adjust crossover probability Pc and the variation probability P m parameter of GA and control evolutionary process, avoid precocious situation; 2. GA optimizes the mode of FNN; Use for reference the thinking of Papadakis and He Suliang etc.; Utilize GA that the major parameter of FL among the FNN and NN is adjusted respectively; Wherein, the parameter of adjustment FL comprises that mainly the subordinate function of fuzzy rule and the major parameter of fuzzy learning rule and adjustment NN comprise study step-length, network weight, hidden layer node numerical value etc.Continuous like this through feedback learning, after stability condition satisfies,, and write down the state parameter of the estimation of this moment with the image implicit expression semantic feature that obtains identification;

(2) image implicit expression semantic feature normalization: for eliminating the influence of external environment factor; Obtain stable image implicit expression semantic feature; This research is further revised with normalization image implicit expression semantic feature and is handled: promptly decide the normal condition comparison according to the state parameter of " the intelligent blackbox model " that picked out with the institute that obtains through sample learning; Its difference is as the input of model; Try to achieve the image implicit expression semantic feature deviation that causes because of environmental difference, thereby obtain normalized image implicit expression semantic feature.

With above-mentioned foundation desirable embodiment of the present invention is enlightenment, and through above-mentioned description, the related work personnel can carry out various change and modification fully in the scope that does not depart from this invention technological thought.The technical scope of this invention is not limited to the content on the instructions, must confirm its technical scope according to the claim scope.

Claims (7)

1. the implicit expression semantic feature method for distilling in aged user's multi-biological characteristic authentication; It is characterized in that: shine upon through aged user's multi-biological characteristic image being carried out semantic analysis of multimode implicit expression and data mining, and extract image implicit expression semantic feature and carry out authentication.

2. the implicit expression semantic feature method for distilling in aged user's multi-biological characteristic according to claim 1 authentication; It is characterized in that: described multi-biological characteristic image comprises people's face, refers to fingerprint and palmmprint more that the concrete steps of the multiple local low-level image feature of its extraction multi-biological characteristic image are following:

A. biometric image pre-service: refer to that by people's face, four emerging system that fingerprint and palmmprint constitute carries out pre-service;

B. extract low-level image feature: extract invariant moment features, Garbor filter characteristic, direction equalization characteristic and half-tone information entropy characteristic;

C. extract the local low-level image feature in the low-level image feature: successively through to the selection of RP, based on the extraction of the ROI of RP and the division of ROI, at last through extracting the local low-level image feature of image.

3. the implicit expression semantic feature method for distilling in aged user's multi-biological characteristic according to claim 1 authentication is characterized in that: the concrete steps of described multimode implicit expression semantic analysis are following:

A. the image array of low-level image feature makes up: adopt multiple low-level image feature, make up the image array of each user's low-level image feature;

B. parallel two-dimentional nonnegative matrix decomposition algorithm: the image array to low-level image feature carries out the diagonalization processing earlier; Again diagonalizable matrix being carried out the row matrix direction decomposes; And then former diagonalizable matrix is carried out transpose process obtain column direction information, the basis matrix that obtains is carried out basis matrix orthogonalization;

C. fuzzy C-means clustering: utilize the fuzzy C-means clustering method in the programming tool case to carry out cluster.

4. the implicit expression semantic feature method for distilling in aged user's multi-biological characteristic according to claim 1 authentication is characterized in that: the concrete steps of the data after the semantic analysis of multimode implicit expression being carried out the data mining mapping are following:

A. beta pruning Algorithm Analysis: the network structure predefine parameter of initialization fuzzy neural network, import first eigenvectors, produce article one fuzzy rule, any input feature value is calculated the distance of itself and first eigenvectors, draw minimum value d _MinThereby, calculate actual output error e _iIf, error e _iGreater than regular importance fault value f, then produce new fuzzy rule, thus the adjustment network architecture parameters;

B. extract image implicit expression semantic feature: comprise off-line learning stage and on-line testing stage.

5. the implicit expression semantic feature method for distilling in aged user's multi-biological characteristic according to claim 4 authentication is characterized in that: the described off-line learning stage is to be used for training fuzzy neural network and to adopt the beta pruning algorithm that the network structure of fuzzy neural network is done dynamic adjustment through learning sample.

6. the implicit expression semantic feature method for distilling in aged user's multi-biological characteristic according to claim 4 authentication; It is characterized in that: the described on-line testing stage is to utilize the fuzzy neural network that trains that test sample book is tested, thereby it is semantic to extract image implicit expression.

7. the implicit expression semantic feature method for distilling in aged user's multi-biological characteristic according to claim 1 authentication is characterized in that: the concrete steps that extraction image implicit expression semantic feature is revised are following:

A. introduce self-adaptation dynamic feedback structure: in intelligent blackbox model, utilize the self-adaptation feedback arrangement of band GA optimized Algorithm to extract the semantic feature and the state parameter of identification based on fuzzy neural network;

B. image implicit expression semantic feature normalization: the state parameter through picking out intelligent blackbox model and with obtain through sample learning decide normal condition and compare; Draw the input of difference as model, thus the image implicit expression semantic feature deviation that must cause because of environmental difference.

Images