CN104063715B - A kind of face classification method based on the nearest feature line - Google Patents

Abstract

The invention discloses a kind of face classification method based on the nearest feature line, with arest neighbors characteristic theory as foundation, define a kind of new weighted index, characteristic line method is simplified after proposing the criterion based on weighted index and improving, construct a kind of face classification device for being suitable for illumination and the various changes of attitude, there are lower computation complexity, less recognition time, and more preferably robustness compared to other graders.The grader extracts the feature of training storehouse image first by principal component analysis, builds training storehouse matrix and extracts sample image feature, builds test sample vector.Then weight coefficient is calculated, and dicision rules, the nearest feature line method that structure is simplified is formulated according to weight coefficient.It is in the case of various test result indicate that, below identical hardware environment, the grader has smaller computational complexity and more preferable robustness compared with other graders.

Description

A kind of face classification method based on the nearest feature line

Technical field

Computer skill is utilized the present invention relates to a kind of face classification method based on the nearest feature line, more particularly to one kind Art, digital image processing techniques, mode identification technology etc. realize the automatic classification of face and sentence method for distinguishing, belongs to biological characteristic In identification field the technology with identification is extracted on face characteristic.

Background technology

1st, face recognition technology

Recognition of face has become an important research direction in living things feature recognition, its key technology be feature to The extraction of amount and the realization of sorting technique.Researcher proposes substantial amounts of face identification method, wherein being carried on characteristic vector Taking popular has principal component analysis (Principal Component Analysis, abbreviation PCA), linear discriminant analysis (Linear Discriminant Analysis, abbreviation LDA) etc.；PCA is a kind of unsupervised algorithm, and it is by solving polynary change The characteristic value of the covariance matrix of amount obtains fundamental component.On sorting technique, popular has K arest neighbors (k- Nearest Neighbor, abbreviation KNN) method, arest neighbors subspace method, SVMs (Support Vector Machine, abbreviation SVM) and grader (Sparse Representation-based based on compressed sensing Classification, abbreviation SRC) etc..

2nd, pivot constituent analysis

Pivot constituent analysis be otherwise known as KL conversion, calculate KL conversion generator matrix ∑, can be the total of training sample Body stroll matrix S_t, or scatter matrix S between the class of training sample_bGenerated by training set Deng, scatter matrix.

Total population scatter matrix can be expressed as：

If taking total population scatter matrix S_tAs generator matrix ∑, note

Then ∑ can be write as：

∑=XX^T∈R^m×m

If by scatter matrix S in class_bAs KL convert generator matrix ∑, i.e.,：

Here c is that training sample concentrates pattern class number,It is that training sample concentrates all kinds of pattern samples Mean value vector, note：

Then generator matrix ∑ is：

∑=XX^T∈R^n×n

The characteristic value and characteristic vector of generator matrix are finally calculated, subspace is constructed, training image and test image are thrown Shadow is to the point corresponded in feature space, in each width image projection to space in subspace.And then can be known with pattern Other theory is classified.

3rd, nearest feature line method

L quasi-modes are provided with, wherein kth class has N_kIndividual sampleTo test sample y and N_kIn individual sample Arbitrary sampleDefine distance：Wherein,Represent y to vector Intersection point, andArrived by calculating yThe distance of characteristic curve, then classification results are as follows, ifThen y belongs to the 1st class.

The content of the invention

Goal of the invention：In order to overcome the deficiencies in the prior art, the present invention to provide a kind of based on the nearest feature line Face identification method, maintain with existing grader identical calculations complexity in the case of, discrimination is higher, robustness is more preferable.

Technical scheme：To achieve the above object, the technical solution adopted by the present invention is：

A kind of face resolving method based on the nearest feature line, comprises the following steps：

(1) training storehouse is set up：Using PCA methods extract sample characteristic value, using extract the characteristic value as training number According to the base vector for obtaining proper subspace, sample is projected to by proper subspace according to the base vector, obtain sample in feature Coordinate in subspace；Set up training storehouse matrix A=[A₁, A₂..., A_k]∈R^m×n, wherein m is each after PCA methods are sampled The dimension of sample, n is the sum for training sample in storehouse, and k is the sum for training sample class in storehouse, A_kFor each class trains picture Set；

(2) by picture projection to be measured to the proper subspace, coordinate of the picture to be measured in proper subspace is obtained

(3) weight coefficient w is calculated_j, and tentatively judged, comprise the following steps：

(31) error function is definedIf

Wherein,It is sample in training storehouse, 1≤j≤n, j is natural number；

(32) local covariance matrix isBeing calculated weight coefficient is

Wherein, 1≤k≤n, k are natural number；1≤l≤n, 1≤m≤n, l, m are natural number；

(33) w is calculated_jThe weight coefficient of each sample class in correspondence training storehouse

(34) weight vectors discriminant index is calculated

(35) threshold tau of weight vectors discriminant index is designed：

Wherein,

(36) size of weight discriminant index W (x) and threshold tau is compared：If W (x) ＞ τ, directly exportMould is maximum Corresponding class is classification results；

(4) if W (x)≤τ, is handled as follows：

(41) amendment training storehouse：Pick outX maximum sample class re-establish train storehouse matrix A '= [A_max1, A_max2..., A_maxX]∈R^m×n；

(42) calculate and simplify characteristic curve：It is by any two images in amendment training storehouseIn feature space To a characteristic curve

(43) coordinate of the picture to be measured in proper subspace to the characteristic curve is calculatedDistanceWherein,

(44) according to describedClassification is obtained

Wherein, N_kIt is the number of samples in each class sample, 1≤k≤n, k_cIt is the number of C class samples.

Beneficial effect：The face resolving method based on the nearest feature line that the present invention is provided, is managed with the nearest feature line By being foundation, a kind of new weight coefficient W (x) is defined, criterion and threshold value calculation method based on W (x) construct one It is individual suitable for illumination and the face classification device of the various changes of attitude, compared to other graders, computation complexity is close, discrimination Higher, robustness is more preferable；In the case of illumination variation and multi-pose Face, the recognition success rate of this method is more than 98%； This method can reach discrimination higher to various features data (such as PCA, LDA, stochastical sampling etc.), in sample characteristics dimension Under conditions of number is less, discrimination higher still can be reached, this feature of the method can reduce sampling request and subtract Few data space, so as to reduce the cost of recognition of face, preferably (such as battery is powered, storage suitable for resource-constrained Capacity is small etc.) hardware environment；In noise jamming and in the case of blocking less than 50%, the method still has preferable Discrimination, has more preferable recognition success rate, more preferable robustness, under adverse circumstances compared to classic algorithms such as NFL, KNN Recognition of face has good adaptability and validity, and has larger raising on computational complexity.

Brief description of the drawings

Fig. 1 is flow chart of the invention；

The effect classified when Fig. 2 is W (x)=0.47；

The effect classified when Fig. 3 is W (x)=0.98；

Fig. 4 is influence of the weight discriminant index to recognition success rate；

Fig. 5 is this algorithm and NFL method comparisons in the case where salt-pepper noise is superimposed；

Fig. 6 is this algorithm and NNL method comparisons in the case where salt-pepper noise is superimposed；

Fig. 7 is this algorithm and KNN method comparisons in the case where salt-pepper noise is superimposed；

Fig. 8 is this algorithm and NFL method comparisons in the case where superposition block is blocked；

Fig. 9 is this algorithm and NNL method comparisons in the case where superposition block is blocked；

Figure 10 is this algorithm and KNN method comparisons in the case where superposition block is blocked.

Specific embodiment

The present invention is further described below in conjunction with the accompanying drawings.

It is as shown in Figure 1 a kind of face resolving method based on the nearest feature line, comprises the following steps：

(1) training storehouse is set up：The characteristic value of sample is extracted using PCA methods, is obtained as training data using the characteristic value of extraction To the base vector of proper subspace, sample is projected to by proper subspace according to base vector, obtain sample in proper subspace Coordinate；Set up training storehouse matrix A=[A₁, A₂..., A_k]∈R^m×n, wherein, m is each sample after the sampling of PCA methods Dimension, n is the sum for training sample in storehouse, and k is the sum for training sample class in storehouse, A_kThe collection of picture is trained for each class Close；

(2) by picture projection to be measured to proper subspace, coordinate of the picture to be measured in proper subspace is obtained

(3) weight coefficient w is calculated_j, and tentatively judged, comprise the following steps：

(31) error function is definedIfWherein,It is sample, 1 in training storehouse ≤ j≤n, j are natural number；

(32) local covariance matrix isBeing calculated weight coefficient isIts In, 1≤k≤n, k is natural number；1≤l≤n, 1≤m≤n, l, m are natural number；

(33) calculate per the corresponding weight coefficient of classIt is weight coefficient w_jEach sample in correspondence training storehouse The coefficient of classification；

(34) weight discriminant index is calculated

(35) judged according to the size for calculating weight discriminant index W (x)：

If W (x)=1, max_i||δ_i(x)||₂/||x||₂=1, illustrate that weight coefficient is only distributed in a class substantially；

If W (x)=0,Illustrate that weight coefficient is almost distributed in each class；

Therefore, it can design distribution situations of the threshold tau ∈ (0,1) for weight discriminant index to represent weight coefficient, Specific design method is：

Wherein,

(36) size of weight discriminant index W (x) and threshold tau is compared：

If W (x) ＞ τ, illustrate that weight coefficient distribution is more concentrated, the effect of classification preferably, can directly export residual error most Small sample class is classification results；

If W (x)≤τ, illustrate that weight coefficient distribution is not good, preferably, classifying quality is bad, it is necessary to reduce instruction for the effect of classification Practice the scope in storehouse, carry out subseries again；

(4) if W (x) is less than τ, it is handled as follows：

(41) amendment training storehouse：Pick outX maximum sample class re-establish train storehouse matrix A '= [A_max1, A_max2..., A_maxX]∈R^m×n；

(42) calculate and simplify characteristic curve：It is by any two images in amendment training storehouseIn feature space To a characteristic curve

(43) coordinate of the picture to be measured in proper subspace to the characteristic curve is calculatedDistanceWherein,

(44) according to describedClassification is obtained

Wherein, N_kIt is the number of samples in each class sample, 1≤k≤n, k_cIt is the number of C class samples.

Just the present invention some detailed problems in the specific implementation are illustrated below.

1st, the test database selected is ORL face databases, UMIST face databases, Extended YaleB faces Database, includes facial image, the mainly change in direction and angle in these three databases.

2nd, feature is extracted using PCA methods, experiment shows, relative to stochastical sampling, PCA has success rate higher.

1) face database is read in, after normalization face database, selects a number of image construction to train everyone in storehouse Collection, remaining constitutes test set.If image is N*M after normalization, pixel is connected by row and constitutes N*M dimensional vectors, can seen Make a point in N*M spaces, can use a lower-dimensional subspace instead by KL changes to describe this image.

2) set in facial image database by N width facial images, be X with vector representation₁, X₂..., X_N, seek its face mean chart As beingThus the inequality of each image is drawn

3) covariance matrix is calculatedThe eigenvalue λ of calculating matrix C_kWith corresponding characteristic vector μ_k.Operand is larger in actual calculating, in order to reduce operand, the inequality of each image is formed into a matrix：X '=[X₁', X₂' ... X_N'], then covariance matrix can be write asIt is theoretical according to linear algebra, X ' (X ') will be calculated^T Eigenvalue λ_jWith corresponding characteristic vector Φ_jProblem be converted to calculating (X ')^TThe eigenvalue λ of X '_jWith corresponding characteristic vector Φ_j', obtain Φ_j' rear Φ_jCan be byObtain.And then the eigenvalue λ of Matrix C is obtained by SVD theorems_k。

4) training image is projected into proper subspace, the inequality of all of N width facial image in face database is empty to this Between project, obtain respective projection vector Y₁, Y₂..., Y_N：

(Y_i) T=[y_1i, y_2i..., y_Mi], i=1,2 ..., N

y_ji=(u_j)^TX′_j, j=1,2 ..., M

Wherein, u_jIt is characterized Vector Groups, X '_jIt is training image inequality.

Composing training matrix A=[Y₁, Y₂..., Y_N], image vector is arranged by class order.

3rd, classify first, for the test pictures for giving, be projected into proper subspace, obtaining property coordinate vector

1) weight coefficient is calculated, and is tentatively judged, define error functionIfWherein,It is training storehouse sample, common n, 1≤j≤n, j is natural number；

2) local covariance matrix isBeing calculated weight coefficient isIts In, 1≤k≤n, k is natural number；1≤l≤n, 1≤m≤n, l, m are natural number；

3) calculate per the corresponding weight coefficient of classIt is each sample class in weight coefficient w correspondence training storehouse Other coefficient, and calculate weight discriminant index

4th, a threshold tau ∈ (0,1) for weight discriminant index is designed to represent the distribution situation of weight coefficient.

5th, designWherein k is the sum for training sample class in storehouse：

If W (x) ＞ τ, illustrate that weight coefficient distribution is more concentrated, directly exportThe maximum corresponding class of mould is classification As a result；

If W (x)≤τ, illustrate that weight coefficient distribution is not good, be then handled as follows, it is necessary to reduce the scope of training pants, enter Capable subseries again；

6th, appropriate selection τ can effectively improve recognition success rate, as shown in Figure 4：

1) pick outX maximum class re-establish a new less training storehouse matrix A '= [A_max1, A_max2, A_max3]；In most cases, correct classification is included in revised less training storehouse, thus Identification range is reduced, X is given by

2) calculate and simplify characteristic curve：It is by any two images in amendment training storehouseObtained in feature space One characteristic curve

3) coordinate of the picture to be measured in proper subspace to the characteristic curve is calculatedDistanceWherein,

4) according to describedClassification is obtained

Wherein, N_kIt is the number of samples in each class sample, 1≤k≤n, k_cIt is the number of C class samples.

The following detailed description of experimental result of the invention：

1st, the database that experiment of the invention is used is international ORL, UMIST, Extended YaleB face numbers According to storehouse.Wherein ORL storehouses, altogether comprising 40 volunteers, everyone contains 10 pictures, and pixel is 92*112, totally 400 figures Piece.We select everyone 5 images as training storehouse, and 5 used as test image in addition.For UMIST storehouses, include altogether Everyone chooses 18 images and uses for 20 volunteers, and pixel is 92*112, wherein 3 used as training image, remaining conduct Test image.For Extended YaleB storehouses, altogether comprising 38 volunteers, everyone chooses 58 images and uses, pixel It is 168*192.

2nd, one is tested：Fig. 4 is displayed in W (x) and takes influence of the different values to successfully tested rate and time, and this experiment exists Carried out on UMIST storehouses.Abscissa represents the value of W (x), from 0 to 1.Left ordinate represents recognition success rate, and right ordinate is represented Time.Experiment shows that with the raising of W (x) values after W (x) ＞ 0.5, this algorithm can obtain higher than NFL algorithm Success rate, and with the raising of W (x), the testing time also increased.

3rd, two are tested：Fig. 5-Figure 10 is the experiment of test image superimposed noise.Experiment is the reality carried out on ORL face databases Test, from randomface, eigenface, fisherface are used as feature extraction mode.Fig. 6 is the reality for being superimposed salt-pepper noise Test, Fig. 7 is to be superimposed the experiment that random block is blocked.Abscissa represents image percentage shared by noise, and ordinate represents identification Success rate.Experiment shows, in the case of superimposed noise, than NFL, KNN, NNL algorithm have preferably identification effect to this paper algorithms Really, especially when selecting eigenface as feature extraction, this algorithm can be obtained in the case of 50% noise accounting The lifting of nearly 20% recognition success rate.

4th, three are tested：Table 1 is tested for algorithm complex, and experiment is carried out on Extended YaleB storehouses, special in have chosen 3 The mode of extraction is levied, is respectively randomface, eigenface, fisherface, this algorithm and classical sorting algorithm are such as KNN, NNL etc. have approximate algorithm complex, but the recognition success rate of algorithm and robustness are all better than classic algorithm.With Other some algorithms such as NFL, NFP, SVM are compared, and the testing time of this algorithm has significantly to reduce.

The operation time of table 1 tests (s)

	Randomfaces	Eigenfaces	Fisherfaces
				SFL	110.88	143.42	188.85
KNN	106.24	112.81	135.06
				NNL	107.95	133.76	175.51
NFL	598.23	639.03	789.35
				NFP	7890.06	9180.31	9862.89
SVM	1824.08	483.81	565.25

The above is only the preferred embodiment of the present invention, it should be pointed out that：For the ordinary skill people of the art For member, under the premise without departing from the principles of the invention, some improvements and modifications can also be made, these improvements and modifications also should It is considered as protection scope of the present invention.

Claims (1)

1. a kind of face resolving method based on the nearest feature line, it is characterised in that：Comprise the following steps：

(1) training storehouse is set up：The characteristic value of sample is extracted using PCA methods, is obtained as training data using the characteristic value of extraction To the base vector of proper subspace, sample is projected to by proper subspace according to the base vector, obtain sample empty in feature Interior coordinate；Set up training storehouse matrix A=[A₁,A₂,...,A_k]∈R^m×n, wherein m is each sample after the sampling of PCA methods Dimension, n be train storehouse in sample sum, k be train storehouse in sample class sum, A_kThe collection of picture is trained for kth class Close；

(2) by picture projection to be measured to the proper subspace, coordinate of the picture to be measured in proper subspace is obtained

(3) weight coefficient w is calculated_j, and tentatively judged, comprise the following steps：

(31) error function is definedIf

Wherein,It is sample in training storehouse, 1≤j≤n, j is natural number；

(32) local covariance matrix isBeing calculated weight coefficient is

Wherein, 1≤z≤n, z are natural number；1≤l≤n, 1≤m≤n, l, m are natural number；

(33) w is calculated_jThe weight coefficient of each sample class in correspondence training storehouseI takes 1~k；

(34) weight vectors discriminant index is calculated

(35) threshold tau of weight vectors discriminant index is designed：

$\mathrm{\τ}=\frac{\sqrt{k}\×(1-X/k)-1}{\sqrt{k}-1}$

Wherein,

(36) size of weight discriminant index W (x) and threshold tau is compared：If W (x) ＞ τ, directly exportThe maximum correspondence of mould Class be classification results；

(4) if W (x)≤τ, is handled as follows：

(41) amendment training storehouse：Pick outX maximum sample class of mould re-establish train storehouse matrix A '=[A_max1, A_max2,...,A_maxX]∈R^m×n；

(42) calculate and simplify characteristic curve：It is by any two images in amendment training storehouseOne is obtained in feature space Characteristic curve

(43) coordinate of the picture to be measured in proper subspace to the characteristic curve is calculatedDistanceWherein,

(44) according to describedClassification is obtained

Wherein, k_cIt is the number of C class samples.