CN103927531A - Human face recognition method based on local binary value and PSO BP neural network - Google Patents

Abstract

The invention discloses a human face recognition method based on a local binary value and a PSO BP neural network. The method comprises the steps that first, all kinds of human face images in a known human face library are divided into a training sample set and a testing sample set in a non-overlapped mode, and normalization and local binary preprocessing are conducted on the images; second, two-dimensional discrete wavelet transformation is conducted on the preprocessed images, the influence of diagonal line component is removed, weight fusion is conducted on other three frequency band components, then two-dimensional discrete cosine transform is conducted on the fused images, and a zigzag scanning mode is used for extracting a main transformation coefficient matrix; then, the initial weight value and the threshold value of the PSO BP neural network are used for conducting network training; at last, the data of the testing sample set are sent to the trained BP neural network for testing, and the recognition rate is calculated. According to the human face recognition method based on the local binary value and the PSO BP neural network, high computing efficiency and high recognition capacity are achieved, and the method is suitable for human face recognition systems.

Description

A kind of face identification method based on local two-value and particle group optimizing BP neural network

Technical field

The invention belongs to a kind of face identification method, relate in particular to a kind of face recognition algorithms based on local two-value and particle group optimizing BP neural network, belong to intelligent mode identification and image processing field.

Background technology

In recent years, face recognition technology has obtained swift and violent development, and the appearance of a large amount of high-performance algorithm makes it from laboratory, move towards commercialization.Yet up to the present, face recognition technology is still faced with huge challenge: (1) illumination, background, attitude, expression, shelter and change of age; (2) image-forming condition and equipment difference; (3) restriction of data scale etc.Therefore, in face recognition technology, high discrimination problem is not still solved.

The variation of ambient light is to affect one of principal element of recognition of face precision.Research finds, the difference of same facial image under different illumination conditions often than different facial images the difference under identical illumination condition much bigger.And existing a lot of face identification system is all to use, there is significant limitation under strict illumination condition.Therefore before face classification identification, illumination pretreatment is necessary, and pretreated quality directly has influence on the quality of system performance.

Facial image is generally two dimensional image, and contained information data amount is huge, easily causes dimension disaster, is unfavorable for Classification and Identification.In order to reduce the complexity of calculating, improve arithmetic speed and the discrimination of system, must compress dimension-reduction treatment to image, by the least possible data representation information as much as possible.

The selection of sorter is the another key factor that affects recognition of face rate.The kind of sorter has a lot, but because neural network can obtain additive method, be difficult to identification rule and stealthy expression of realizing, and its parallel processing mode can significantly improve computing velocity, therefore, enjoy vast focus of attention, be widely applied in every field.Although neural network has reduced the complexity of computing to a certain extent, but the setting of each parameter of neural network does not have specific principle algorithm, need us by virtue of experience to carry out value, in addition, in the speed of convergence of neural network be very easily absorbed in aspect the problems such as local minimum and also thoroughly do not solve.

Summary of the invention

Goal of the invention: in order to overcome existing deficiency, the present invention proposes a kind of face identification method based on local two-value and particle group optimizing BP neural network.

Technical scheme: a kind of face identification method based on local two-value and particle group optimizing BP neural network, comprises the following steps:

(1) every class facial image in known face database is randomly drawed to some as training sample set I _train={ X ₁, X ₂..., X _j..., X _a, wherein, X _jfor each training sample image, A is number of training;

(2) to training sample, concentrate the gray level image of every width M * N pixel to carry out geometrical normalization processing, be normalized to the image of H * H size, be designated as I ' _train, 0 < H≤min (M, N) wherein;

(3) the training set image I after utilizing local two value-based algorithms to geometrical normalization ' _trainextract illumination invariant, remove illumination effect, obtain the training set image I after photo-irradiation treatment " _train;

(4) to the image set I after local two-value algorithm process in step (3) " _trainbe weighted two-dimensional discrete wavelet conversion, the image set I after being converted _{train, DWT};

(5) image set I step (4) being obtained _{train, DWT}in each sample image do two-dimension discrete cosine transform, obtain transform coefficient matrix Y={Y ₁, Y ₂..., Y _h..., Y _a, wherein, Y _hconversion coefficient vector for each sample image obtains after two-dimension discrete cosine transform, then, utilizes zigzag scan mode to launch each vector in transform coefficient matrix Y, last, extracts each and launches vectorial principal component, forms optimal characteristics vector E;

(6) neural network parameter is set, determines the topological structure of BP neural network, input layer is counted Q, hidden layer node is counted W, output layer nodes Z, activation function Sigmoid function;

(7) by particle cluster algorithm Optimized BP Neural Network weights and bias;

(8) global optimum step (7) being obtained is mapped as initial weight and the threshold value of neural network, training BP neural network;

(9) using remaining image in face database as test sample book collection I _test, the processing by its repeating step (2) to step (5), is then input to test sample book data in the resulting BP neural network having trained of step (8) and tests, and calculates discrimination.

Described step (3) is specially:

First, by I ' _trainin every width image be divided into the fritter of n * n, in each fritter, the average of n * n pixel gray-scale value, as a pixel value of new images, obtains piecemeal training set image I _{train_block};

Then, utilize local two value-based algorithms to describe the illumination invariant measure feature of facial image, for I _{train_block}in any image I, I ∈ I _{train_block}, its any point (x _c, y _c) and the local two-value feature operator LBP that is evenly distributed on this central point P neighborhood point be around:

$\left\{\begin{array}{c}{\mathrm{LBP}}_{R,P}=\underset{q=0}{\overset{P-1}{\mathrm{\&Sigma;}}}s(g_q-g_c)\&CenterDot;2^q,\\ s\left(x\right)=\left\{\begin{array}{c}1,x\&GreaterEqual;0\\ 0,x<0\end{array}.\right.\end{array}\right.---\left(1\right)$

Wherein, g _crepresent central pixel point (x in region _c, y _c) gray-scale value, g _q(q=0,1 ..., P-1) represent to be evenly distributed on central point (x _c, y _c) be the center of circle, the gray-scale value of P sampled point on the circumference that radius is R.

Described step (4) is specially:

At image set I " _trainin choose any piece image F, it is carried out to two-dimensional discrete wavelet conversion, obtain LL, LH, HL, the image on HH four direction, is designated as F _lL, F _lH, F _hL, F _hH, the image F ' after weighting is:

F′＝a ₀F _LL+a ₁F _LH+a ₂F _HL (2)

Wherein, a ₀, a ₁, a ₂for weighting coefficient, constraint condition is a ₀+ a ₁+ a ₂=1; Low frequency component F _lLfor the smoothed image of original image, kept the most information of original image; F _lHcomponent has kept the vertical edge details of original image; F _hLcomponent has kept the horizontal edge details of original image; F _hHcomponent has kept the edge details in original image diagonal.

Described step (7) is specially:

A., population parameter is set, comprises particle population scale B, dimension D, maximum iteration time T _max, study factor c ₁and c ₂, inertia weight ω, maximal rate υ _max, maximum position x _max, anticipation error minimum value ε, random initial velocity and the position that produces particle in allowed band;

B. calculate the fitness J of each particle _a;

C. according to fitness J _adetermine the optimum extreme value of individuality and global optimum's extreme value of each particle, by current fitness value and the comparison of the historical optimal-adaptive degree of each particle, the little person of fitness is as the optimum extreme value P of individuality _best, with the comparison of whole population optimal-adaptive degree, the little person of fitness is as the extreme value G of global optimum _best;

D. upgrade speed and the position of each particle, and the speed after consider upgrading and position are whether in limited range;

${\mathrm{\&upsi;}}_{\mathrm{id}}^{T+1}={\mathrm{\&omega;\&upsi;}}_{\mathrm{id}}^T+c_1{r}_1(p_{\mathrm{id}}^T-x_{\mathrm{id}}^T)+c_2{r}_2(p_{\mathrm{gd}}^T-x_{\mathrm{id}}^T)---\left(3\right)$

$x_{\mathrm{id}}^{T+1}=x_{\mathrm{id}}^T+{\mathrm{\&upsi;}}_{\mathrm{id}}^T---\left(4\right)$

Wherein, i=1,2 ..., B, d=1,2 ..., D, it is the d dimension component of the T time iteration particle i velocity; it is the d dimension component of the T time iteration particle i position vector; for the individual current desired positions P of particle i _bestd dimension component; for the current desired positions G of colony _bestd dimension component; r ₁and r ₂for obeying [0,1] equally distributed random number;

Constraint condition 1: if ${\mathrm{\&upsi;}}_{\mathrm{id}}^{T+1}>{\mathrm{\&upsi;}}_{\max },$ ${\mathrm{\&upsi;}}_{\mathrm{id}}^{T+1}={\mathrm{\&upsi;}}_{\max };$ If ${\mathrm{\&upsi;}}_{\mathrm{id}}^{T+1}<-{\mathrm{\&upsi;}}_{\max },$ ${\mathrm{\&upsi;}}_{\mathrm{id}}^{T+1}=-{\mathrm{\&upsi;}}_{\max };$

Constraint condition 2: if $x_{\mathrm{id}}^{T+1}>x_{\max },$ $x_{\mathrm{id}}^{T+1}=x_{\max };$ If $x_{\mathrm{id}}^{T+1}<-x_{\max },$ $x_{\mathrm{id}}^{T+1}=-x_{\max };$

E. relatively whether iterations reaches maximum T _maxor whether square error reaches precision ε requirement, if meet, algorithm convergence, records the individual optimal value P of last iteration _bestwith global optimum G _best; Otherwise, return to step c.

The computing method of described step b are specially:

$J_a=\frac{1}{2A}\underset{k=1}{\overset{A}{\mathrm{\&Sigma;}}}\underset{t=1}{\overset{Z}{\mathrm{\&Sigma;}}}{(y_k^t-c_t^k)}^2---\left(5\right)$

Wherein, a=1,2 ..., B; it is the desired output of t network output neuron of k sample; it is the real output value of t network output neuron of k sample.

Principle of work: first the present invention by non-overlapping training sample set and the test sample book collection of being divided into of every class facial image in known face database, enters normalization and local two-value pre-service to image; Secondly, pretreated image is done to two-dimensional discrete wavelet conversion, remove the impact of diagonal line component, by its excess-three band component Weighted Fusion, again the image after merging is done to two-dimension discrete cosine transform, utilize zigzag scan mode to extract its main transform coefficient matrix; Again, utilize the initial weight of particle cluster algorithm Optimized BP Neural Network and threshold value to carry out network training; Finally, test sample book collection data are sent in the BP neural network having trained and are tested, calculate discrimination.

Beneficial effect: the present invention removes illumination effect by local two value-based algorithms, weighting wavelet transform combines to carry out feature extraction with discrete cosine transform, weights and the threshold value of recycling particle group optimizing neural network are carried out Classification and Identification, have stronger robustness and optimizing ability.Compared with prior art, the present invention has higher operation efficiency and recognition capability, is applicable to face identification system.

Accompanying drawing explanation

Fig. 1 is block diagram of the present invention;

Fig. 2 is the process flow diagram of particle group optimizing BP neural network in the present invention.

Embodiment

As shown in Figure 1, 2, a kind of face identification method based on local two-value and particle group optimizing BP neural network comprises the following steps:

Step 1: every class facial image in known face database is randomly drawed to some as training sample set I _train={ X ₁, X ₂..., X _j..., X _a, wherein, X _jfor each training sample image, A is number of training, and all the other are as test sample book collection I _test;

Step 2: the gray level image to every width M * N pixel in training set carries out geometrical normalization processing, is normalized to the image of H * H size, is designated as I ' _train, 0 < H≤min (M, N) wherein;

Step 3: the training set image I after utilizing local two value-based algorithms to normalization ' _trainextract illumination invariant, remove illumination effect, obtain the training set image I after photo-irradiation treatment " _train, its process is:

(1) first, by I ' _trainin every width image be divided into the fritter of n * n, in each fritter, the average of n * n pixel gray-scale value, as a pixel value of new images, obtains piecemeal training set image I _{train_block};

(2) utilize local two value-based algorithms to describe the illumination invariant measure feature of facial image, for I _{train_block}in any image I, I ∈ I _{train_block}, its any point (x _c, y _c) and the local two-value feature operator LBP that is evenly distributed on this central point P neighborhood point be around expressed as:

$\left\{\begin{array}{c}{\mathrm{LBP}}_{R,P}=\underset{q=0}{\overset{P-1}{\mathrm{\&Sigma;}}}s(g_q-g_c)\&CenterDot;2^q,\\ s\left(x\right)=\left\{\begin{array}{c}1,x\&GreaterEqual;0\\ 0,x<0\end{array}.\right.\end{array}\right.---\left(1\right)$

Wherein, g _crepresent central pixel point (x in region _c, y _c) gray-scale value, g _q(q=0,1 ..., P-1) represent to be evenly distributed on central point (x _c, y _c) be the center of circle, the gray-scale value of P sampled point on the circumference that radius is R;

Step 4: to above-mentioned image set I after local two-value is processed " _trainbe weighted two-dimensional discrete wavelet conversion, the image set I after being converted _{train, DWT}:

At image set I " _trainin choose any piece image F, it is carried out to two-dimensional discrete wavelet conversion, obtain LL, LH, HL, the image on HH four direction, is designated as F _lL, F _lH, F _hL, F _hH, the image F ' after weighting is:

F′＝a ₀F _LL+a ₁F _LH+a ₂F _HL (2)

Wherein, a ₀, a ₁, a ₂for weighting coefficient, constraint condition is a ₀+ a ₁+ a ₂=1; Low frequency component F _lLfor the smoothed image of original image, kept the most information of original image; F _lHcomponent has kept the vertical edge details of original image; F _hLcomponent has kept the horizontal edge details of original image; F _hHcomponent has kept the edge details in original image diagonal; Because facial image is nonrigid, the less stable of diagonal information, Noise is more, is unfavorable for very much feature extraction, so cast out;

Step 5: to image set I _{train, DWT}in each sample image do two-dimension discrete cosine transform, obtain transform coefficient matrix Y={Y ₁, Y ₂..., Y _h..., Y _a, wherein, Y _hthe conversion coefficient vector obtaining after two-dimension discrete cosine transform for each sample image; Then, each vector in transform coefficient matrix Y is utilized zigzag scan mode launch; Finally, extract each and launch vectorial principal component, form optimal characteristics vector E, be the input of neural network;

Step 6: as shown in Figure 2: neural network parameter is set, determines the topological structure of BP neural network: input layer is counted Q, hidden layer node is counted W, output layer nodes Z; Activation function Sigmoid function;

Step 7: population parameter is set: particle population scale B, wherein 20≤B≤100; Dimension D, wherein D=QW+WZ+W+Z; Maximum iteration time T _max; Study factor c ₁and c ₂, 1≤c wherein ₁≤ 2,1≤c ₂≤ 2, generally get c ₁=c ₂; Inertia weight ω, wherein 0 < ω < 1; Maximal rate υ _max; Maximum position x _max; Random initial velocity, the position that produces particle in allowed band; Anticipation error minimum value ε;

Step 8: the fitness that calculates each particle: first input a particle, calculate all sample standard deviation variances, i.e. the fitness of this particle, that is:

$J_a=\frac{1}{2A}\underset{k=1}{\overset{A}{\mathrm{\&Sigma;}}}\underset{t=1}{\overset{Z}{\mathrm{\&Sigma;}}}{(y_k^t-c_t^k)}^2---\left(3\right)$

Wherein, a=1,2 ..., B; it is the desired output of t network output neuron of k sample; it is the real output value of t network output neuron of k sample; In like manner, continue other particle of input, until calculate the fitness of all particles;

Step 9: according to fitness J _adetermine the optimum extreme value of individuality and global optimum's extreme value of each particle, by current fitness value and the comparison of the historical optimal-adaptive degree of each particle, the little person of fitness is as the optimum extreme value P of individuality _best, with the comparison of whole population optimal-adaptive degree, the little person of fitness is as the extreme value G of global optimum _best;

Step 10: upgrade speed and the position of each particle, and the speed after consider upgrading and position are whether in limited range;

${\mathrm{\&upsi;}}_{\mathrm{id}}^{T+1}={\mathrm{\&omega;\&upsi;}}_{\mathrm{id}}^T+c_1{r}_1(p_{\mathrm{id}}^T-x_{\mathrm{id}}^T)+c_2{r}_2(p_{\mathrm{gd}}^T-x_{\mathrm{id}}^T)---\left(4\right)$

$x_{\mathrm{id}}^{T+1}=x_{\mathrm{id}}^T+{\mathrm{\&upsi;}}_{\mathrm{id}}^T---\left(5\right)$

Wherein, i=1,2 ..., B, d=1,2 ..., D, it is the d dimension component of the T time iteration particle i velocity; it is the d dimension component of the T time iteration particle i position vector; for the individual current desired positions P of particle i _bestd dimension component; for the current desired positions G of colony _bestd dimension component; r ₁and r ₂for obeying [0,1] equally distributed random number;

Constraint condition 1: if ${\mathrm{\&upsi;}}_{\mathrm{id}}^{T+1}>{\mathrm{\&upsi;}}_{\max },$ ${\mathrm{\&upsi;}}_{\mathrm{id}}^{T+1}={\mathrm{\&upsi;}}_{\max };$ If ${\mathrm{\&upsi;}}_{\mathrm{id}}^{T+1}<-{\mathrm{\&upsi;}}_{\max },$ ${\mathrm{\&upsi;}}_{\mathrm{id}}^{T+1}=-{\mathrm{\&upsi;}}_{\max };$

Constraint condition 2: if $x_{\mathrm{id}}^{T+1}>x_{\max },$ $x_{\mathrm{id}}^{T+1}=x_{\max };$ If $x_{\mathrm{id}}^{T+1}<-x_{\max },$ $x_{\mathrm{id}}^{T+1}>x_{\max },$

Step 11: relatively whether iterations reaches maximum T _maxor whether square error reaches precision ε requirement, if meet, algorithm convergence, records the individual optimal value P of last iteration _bestwith global optimum G _best; Otherwise, return to step 9, continue iteration;

Step 12: global optimum is mapped as to initial weight and the threshold value of neural network, training network;

Step 13: by test set I _testrepeating step 2, to the processing of step 5, is input to test sample book data in the BP neural network having trained and tests, and according to the Output rusults of BP neural network, calculates discrimination.

Claims (5)

1. the face identification method based on local two-value and particle group optimizing BP neural network, is characterized in that, comprises the following steps:

(1) every class facial image in known face database is randomly drawed to some as training sample set I _train={ X ₁, X ₂..., X _j..., X _a, wherein, X _jfor each training sample image, A is number of training;

(2) to training sample, concentrate the gray level image of every width M * N pixel to carry out geometrical normalization processing, be normalized to the image of H * H size, be designated as I ' _train, 0 < H≤min (M, N) wherein;

(3) the training set image I after utilizing local two value-based algorithms to geometrical normalization ' _trainextract illumination invariant, remove illumination effect, obtain the training set image I after photo-irradiation treatment " _train;

(4) to the image set I after local two-value algorithm process in step (3) " _trainbe weighted two-dimensional discrete wavelet conversion, the image set I after being converted _{train, DWT};

(5) image set I step (4) being obtained _{train, DWT}in each sample image do two-dimension discrete cosine transform, obtain transform coefficient matrix Y={Y ₁, Y ₂..., Y _h..., Y _a, wherein, Y _hconversion coefficient vector for each sample image obtains after two-dimension discrete cosine transform, then, utilizes zigzag scan mode to launch each vector in transform coefficient matrix Y, last, extracts each and launches vectorial principal component, forms optimal characteristics vector E;

(6) neural network parameter is set, determines the topological structure of BP neural network, input layer is counted Q, hidden layer node is counted W, output layer nodes Z, activation function Sigmoid function;

(7) by particle cluster algorithm Optimized BP Neural Network weights and bias;

(8) global optimum step (7) being obtained is mapped as initial weight and the threshold value of neural network, training BP neural network;

(9) using remaining image in face database as test sample book collection I _test, the processing by its repeating step (2) to step (5), is then input to test sample book data in the resulting BP neural network having trained of step (8) and tests, and calculates discrimination.

2. a kind of face identification method based on local two-value and particle group optimizing BP neural network according to claim 1, is characterized in that, described step (3) is specially:

First, by I ' _trainin every width image be divided into the fritter of n * n, in each fritter, the average of n * n pixel gray-scale value, as a pixel value of new images, obtains piecemeal training set image I _{train_block};

Then, utilize local two value-based algorithms to describe the illumination invariant measure feature of facial image, for I _{train_block}in any image I, I ∈ I _{train_block}, its any point (x _c, y _c) and the local two-value feature operator LBP that is evenly distributed on this central point P neighborhood point be around:

$\left\{\begin{array}{c}{\mathrm{LBP}}_{R,P}=\underset{q=0}{\overset{P-1}{\mathrm{\&Sigma;}}}s(g_q-g_c)\&CenterDot;2^q,\\ s\left(x\right)=\left\{\begin{array}{c}1,x\&GreaterEqual;0\\ 0,x<0\end{array}.\right.\end{array}\right.---\left(1\right)$

Wherein, g _crepresent central pixel point (x in region _c, y _c) gray-scale value, g _q(q=0,1 ..., P-1) represent to be evenly distributed on central point (x _c, y _c) be the center of circle, the gray-scale value of P sampled point on the circumference that radius is R.

3. a kind of face identification method based on local two-value and particle group optimizing BP neural network according to claim 1, is characterized in that, described step (4) is specially:

At image set I " _trainin choose any piece image F, it is carried out to two-dimensional discrete wavelet conversion, obtain LL, LH, HL, the image on HH four direction, is designated as F _lL, F _lH, F _hL, F _hH, the image F ' after weighting is:

F′＝a ₀F _LL+a ₁F _LH+a ₂F _HL (2)

Wherein, a ₀, a ₁, a ₂for weighting coefficient, constraint condition is a ₀+ a ₁+ a ₂=1; Low frequency component F _lLfor the smoothed image of original image, kept the most information of original image; F _lHcomponent has kept the vertical edge details of original image; F _hLcomponent has kept the horizontal edge details of original image; F _hHcomponent has kept the edge details in original image diagonal.

4. according to a kind of face identification method based on local two-value and particle group optimizing BP neural network claimed in claim 1, it is characterized in that, described step (7) is specially:

A., population parameter is set, comprises particle population scale B, dimension D, maximum iteration time T _max, study factor c ₁and c ₂, inertia weight ω, maximal rate υ _max, maximum position x _max, anticipation error minimum value ε, random initial velocity and the position that produces particle in allowed band;

B. calculate the fitness J of each particle _a;

C. according to fitness J _adetermine the optimum extreme value of individuality and global optimum's extreme value of each particle, by current fitness value and the comparison of the historical optimal-adaptive degree of each particle, the little person of fitness is as the optimum extreme value P of individuality _best, with the comparison of whole population optimal-adaptive degree, the little person of fitness is as the extreme value G of global optimum _best;

D. upgrade speed and the position of each particle, and the speed after consider upgrading and position are whether in limited range;

${\mathrm{\&upsi;}}_{\mathrm{id}}^{T+1}={\mathrm{\&omega;\&upsi;}}_{\mathrm{id}}^T+c_1{r}_1(p_{\mathrm{id}}^T-x_{\mathrm{id}}^T)+c_2{r}_2(p_{\mathrm{gd}}^T-x_{\mathrm{id}}^T)---\left(3\right)$

$x_{\mathrm{id}}^{T+1}=x_{\mathrm{id}}^T+{\mathrm{\&upsi;}}_{\mathrm{id}}^T---\left(4\right)$

Wherein, i=1,2 ..., B, d=1,2 ..., D, it is the d dimension component of the T time iteration particle i velocity; it is the d dimension component of the T time iteration particle i position vector; for the individual current desired positions P of particle i _bestd dimension component; for the current desired positions G of colony _bestd dimension component; r ₁and r ₂for obeying [0,1] equally distributed random number;

Constraint condition 1: if ${\mathrm{\&upsi;}}_{\mathrm{id}}^{T+1}>{\mathrm{\&upsi;}}_{\max },$ ${\mathrm{\&upsi;}}_{\mathrm{id}}^{T+1}={\mathrm{\&upsi;}}_{\max };$ If ${\mathrm{\&upsi;}}_{\mathrm{id}}^{T+1}<-{\mathrm{\&upsi;}}_{\max },$ ${\mathrm{\&upsi;}}_{\mathrm{id}}^{T+1}=-{\mathrm{\&upsi;}}_{\max };$

Constraint condition 2: if $x_{\mathrm{id}}^{T+1}>x_{\max },$ $x_{\mathrm{id}}^{T+1}=x_{\max };$ If $x_{\mathrm{id}}^{T+1}<-x_{\max },$ $x_{\mathrm{id}}^{T+1}=-x_{\max };$

E. relatively whether iterations reaches maximum T _maxor whether square error reaches precision ε requirement, if meet, algorithm convergence, records the individual optimal value P of last iteration _bestwith global optimum G _best; Otherwise, return to step c.

5. a kind of face identification method based on local two-value and particle group optimizing BP neural network according to claim 4, is characterized in that, the computing method of described step b are specially:

$J_a=\frac{1}{2A}\underset{k=1}{\overset{A}{\mathrm{\&Sigma;}}}\underset{t=1}{\overset{Z}{\mathrm{\&Sigma;}}}{(y_k^t-c_t^k)}^2---\left(5\right)$

Wherein, a=1,2 ..., B; it is the desired output of t network output neuron of k sample; it is the real output value of t network output neuron of k sample.