CN104239863A - Face recognition method based on simplified blood flow model and improved Weber local descriptor - Google Patents

Abstract

The invention relates to a face recognition method based on a simplified blood flow model and an improved Weber local descriptor. A thermal infrared facial feature having robustness is extracted through the simplified blood flow model and the improved Weber local descriptor to achieve thermal infrared face image recognition. The face recognition method utilizes the simplified blood flow model to convert a thermal infrared face thermogram into a rheogram, and a stable face biological feature can be obtained. Local features of the rheogram are extracted by using the improved Weber local descriptor, and local authentication information can be fully utilized. The features are classified through PCA-method dimensionality reduction by means of a three-order neighbor algorithm, and thermal infrared face image recognition is achieved.

Description

Based on the face identification method simplifying flow model and improvement weber local description

Technical field

The invention belongs to technical field of biometric identification, relate to the identification of thermal infrared facial image, especially a kind of face identification method based on simplifying flow model and improvement weber local description.

Background technology

Infrared face recognition is divided near infrared recognition of face and far infrared recognition of face.The identifying of near infrared recognition of face installs a near-infrared luminous diode on cameras and long logical optical filter obtains a face direct picture clearly, can reduce the impact of ambient lighting on the human face photo obtained to a great extent.But its major defect is the cooperation needing tester, near infrared recognition of face is made not have the advantage of recognition of face passivity.In addition, because near infrared facial image is obtained by the reflected light imaging of face, if the face of tester does not face camera, will hypographously exist, the performance that impact identifies.Compared with near infrared facial image, the imaging of far infrared face is that the heat radiation sent by obtaining face carrys out imaging, is determined by the Temperature Distribution of face, does not need the cooperation of tester.Far infrared recognition of face is also referred to as thermal infrared recognition of face.The present invention is based on far recognition of face, conveniently, below referred to as infrared face recognition.

Infrared face recognition technology is proposed the earliest in 1992 by doctor Prokoski of Mikos company of the U.S..Relative to the face recognition technology of visible ray, under the conditions such as infrared face recognition technology can change at illumination condition, human face posture change, human face expression, cosmetic, photo swindle, obtain better discrimination.Infrared face recognition technology based on the temperature information of face, reflection be human skin's temperature profile, be easily subject to the impact of the factors such as environment temperature.So how eliminating the impact of these factors, extract the infrared face feature of robustness, is the important directions of infrared face recognition research.

Face thermogram is determined by the infrared emanation of face tissue and structure (as vessel size and vascular distribution etc.), although thermogram easily affects by the factor such as environment temperature, mood, but everyone vascular distribution is unique, not reproducible, and this characteristic does not change with the growth at age, so they are relevant with the physiological structure of people as fingerprint, there is uniqueness.So how utilizing the physiological characteristic of vascular distribution to carry out recognition of face, is the important directions improving infrared face recognition performance.

Feature extracting method is the key of recognition of face, and along with the development of infrared face recognition technology, various feature extracting methods is also suggested.Based on the infrared face recognition method of local binary pattern (LBP), LBP method can extract abundant local grain information, efficiently must describe texture.Weber local description (WLD) is a kind of new Local Feature Extraction.The difference of WLD and LBP is: the reflection of LBP method be difference between center pixel and adjacent pixels, and WLD method had both reflected the difference between center pixel and adjacent pixels, further comprises the information of gradient direction change.But WLD is Shortcomings in compute gradient directional information, the calculating of gradient direction is easily subject to the interference of picture noise, affects the effect of final thermal infrared recognition of face.

Summary of the invention

The object of the invention is to overcome prior art deficiency, provide a kind of distinguishing ability utilizing image processing techniques and intellectual technology to solve thermal infrared recognition of face problem stronger based on simplifying flow model and improving the face identification method of weber local description.

The technical solution used in the present invention is:

A kind of thermal infrared face identification method of weber local description based on simplifying flow model and improvement, comprise a spinning machine, this spinning machine top is provided with main shaft and lower main axis, upper main shaft is lifted on upper top in spinning machine and passes through air cylinder driven, jig is fixed with bottom upper main shaft, at the spinning machine inner bottom part with lower main axis, lower main axis is installed, spinning machine bottom shell is run through and by Driven by Hydraulic Cylinder bottom this lower main axis, the mould placing wheel blanks is fixed with at the lower main axis top corresponding with jig, this spinning machine front end is open and spinning machine is interior in three madial walls, corresponding with mould three madial wall positions are separately installed with three spray assemblies.

And, each spray assembly comprises substrate, grip block, roller, solenoid valve, atomization storehouse and shower nozzle, substrate is arranged on the position of the corresponding spray assembly of three madial walls respectively, each substrate is all installed the grip block of pair of parallel setting, one roller is installed between two grip blocks, grip block keeps tilting relative to mould, ensure that the wheel blanks outside surface on roller and mould leaves minim gap, on each substrate, end face is provided with two solenoid valves, two solenoid valves control a compressed air conveying pipeline and a release agent transfer pipeline respectively, compressed air conveying pipeline entrance point is connected with compressed air source, compressed air conveying pipeline endpiece is connected with atomization storehouse import, release agent transfer pipeline entrance point is connected with release agent charging basket, release agent transfer pipeline entrance point endpiece is connected with atomization storehouse import, the outlet of atomization storehouse is connected with shower nozzle.

And, compressed air conveying pipeline endpiece with the pipeline of atomization storehouse import are provided with air inflow variable valve.

Advantage of the present invention and good effect are:

By simplifying flow model, the temperature profile information (thermogram) thermal infrared face being subject to outside environmental elements impact is converted into the biological information (rheography) of human body, highlight the feature of face, weaken the non-characteristic informations such as contextual factor to the impact identified.Rheography not only reflects the feature of thermogram, but also reflects the face physiological characteristic organized between artery and skin surface; By the weber Feature Descriptor improved, utilize differential excition and calculate the differential excition information and Gradient direction information that calculate respectively with tropism Sobel operator, to noise, there is stronger antijamming capability, build more stable thermal infrared face characteristic histogram, obtain the thermal infrared face characteristic of robustness more, improve thermal infrared recognition of face performance.

Accompanying drawing illustrates:

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

Fig. 2 is the contrast of thermogram and rheography;

Fig. 3 is the process flow diagram improving weber local description.

Embodiment:

Below by accompanying drawing, the invention will be further described in conjunction with specific embodiments, and following examples are descriptive, is not determinate, can not limit protection scope of the present invention with this.

Based on the face identification method simplifying flow model and improvement weber local description, it totally as shown in Figure 1, comprises the steps:

(1) thermogram conversion rheography: thermogram is converted to rheography and is realized by the flow model simplified, obtain the blood-stream image of face;

(2) Gaussian smoothing: carry out Gaussian smoothing operation to above-mentioned blood-stream image, removes the interference of noise, improves infrared face recognition performance;

(3) the weber local description improved extracts local feature: carry out local shape factor by the weber local description improved to the rheography of piecemeal, wherein the calculating of directional information utilizes with tropism Sobel operator, obtain more stable local authentication information, the WLD be improved, (1) the WLD (3) completed respectively in training image to step by step is improved and the WLD of test pattern improves, and obtains feature database image information and test pattern information respectively;

(4) PCA method dimensionality reduction: respectively dimensionality reduction is carried out to the WLD that the WLD improved in training image and test pattern improve;

(5) identify: the distance in the lower dimensional space calculating the feature database image information after above-mentioned dimensionality reduction and test pattern information between image, will obtain final recognition result by three rank nearest neighbour classifications.

Each step is described as follows:

Step (1) in thermogram conversion rheography be performance in order to improve infrared face recognition system and stability, first face simplifies flow model, the temperature information (thermogram) being subject to external environment influence is transferred to the physiologic information (rheography) of human body, the kinematic parameter of such as blood flow.And the blood circulation information of human body is stable comparatively speaking, by external environment influence, different people can be identified based on these specific biological informations.

Simplify flow model as follows:

$\mathrm{\ω}=\frac{\mathrm{\ϵ\σ}(T^4-{T_e}^4)}{\mathrm{\α}{c}_b(T_a-T)}$

Wherein, the implication of each physical descriptor is: T _afor artery temperature, T _efor environment temperature, T are skin temperature, ε, σ, α, c _bfor constant parameter.

By simplifying the conversion of flow model, obtain the rheography of face, as shown in Figure 2.

Step (2) in Gaussian smoothing adopt following formula calculate:

I'＝I*G(x,y,σ)

Wherein, * represents convolution algorithm, and gaussian kernel function is as follows:

$G(x,y,\mathrm{\σ})=\frac{1}{2\mathrm{\π}{\mathrm{\σ}}^2}\exp (-\frac{x^2+y^2}{2{\mathrm{\σ}}^2})$

Wherein σ is the variance of Gaussian filter, and when σ value is a bit larger tham 1, final discrimination is higher.

Step (3) in the weber local description of improvement extract before local feature, first piecemeal is carried out to facial image, the object of piecemeal is to extract face local feature better, face is divided into i × j sub-image area, and i represents the number of vertical direction sub-block, and j represents the sub-block number of horizontal direction, such as, the image of one 100 × 80, is divided into 5 × 4 sub-image areas, namely represents that the size of subimage is 20 × 20.; Then utilize the weber local description improved to extract local feature to each subimage respectively, with 3 × 3 operators provided in figure, filtering operation is carried out to subimage, the pixel differential excition value of computed image.The weber local description improved as shown in Figure 3.

The difference information of differential excition Zhi Shi computing center's pixel and surrounding pixel.By the differential excition value sum of surrounding pixel and center pixel than upper center pixel value, obtain a ratio, therefrom can obtain the marked change information of topography, calculate surrounding pixel and the difference sum of center pixel and the ratio of center pixel and adopt following formula:

$G_{\mathrm{ratio}}\left(x_c\right)=\frac{v_s^{00}}{v_s^{01}}$

Wherein v ₀₀and v ₀₁differential excition operator f respectively ₀₀and f ₀₁output, v ₀₀represent the differential excition value sum of surrounding pixel and center pixel, v ₀₁represent original image, x _cit is center pixel.

In order to avoid violent change occurs suddenly ratio, then do the differential excition that arctan (arc tangent) conversion obtains Current central pixel:

$\mathrm{\ξ}\left(x_c\right)=\arctan \left[\frac{v_s^{00}}{v_s^{01}}\right]=\arctan \left[\underset{i=0}{\overset{p-1}{\mathrm{\Σ}}}\left(\frac{x_i-x_c}{x_c}\right)\right]$

This differential excition ξ (x _c) variation range exist in order to form the histogram that can add up, M interval is quantized in differential excition, ξ _m(m=0,1 ... M-1).For each interval ξ _m, have ξ _m=[d _m, u _m], wherein lower interval d _m=(m/M-1/2) π, upper interval u _m=[(m+1)/M-1/2] π.

Directional information described by direction, the graded ratio of image level direction and vertical direction.In order to overcome the shortcoming and defect of original WLD method on gradient direction calculates, introduce isotropy Sobel operator and replace original method to carry out compute gradient directional information, the computing method of the gradient direction of isotropy Sobel operator are as follows:

$\mathrm{\θ}\left(x_c\right)=\arctan \left(\frac{f_s^{11}}{f_s^{10}}\right)$

$f_s^{10}=f_{10}*I,$ $f_s^{11}=f_{11}*I$

Direction θ (x _c) variation range be the same with differential excition, in order to better set up histogram, T principal direction is quantized in the change in direction.Do before a quantization and map f as follows:

${\mathrm{\θ}}^{\′}=\arctan 2(v_s^{11},v_s^{10})+\mathrm{\π},$

$\arctan 2(v_s^{11},v_s^{10})=\left\{\begin{array}{cccc}\mathrm{\&theta;},& v_s^{11}>0& \mathrm{and}& v_s^{10}>0,\\ \mathrm{\&pi;}+\mathrm{\&theta;},& v_s^{11}>0& \mathrm{and}& v_s^{10}<0,\\ \mathrm{\&theta;}-\mathrm{\&pi;},& v_s^{11}<0& \mathrm{and}& v_s^{10}<0,\\ \mathrm{\&theta;},& v_s^{11}<0& \mathrm{and}& v_s^{10}>0,\end{array}\right.$

The expanded range of direction change is to [0,2 π].Quantization function is as follows:

After obtaining differential excition and direction, build the histogram of a 2D, { WLD (ξ _m, Φ _t); M=0,1 ..., M-1; T=0,1 ..., T-1; M and T represents the number in differential excition interval and the number of principal direction respectively.So the 2D histogram obtained, every a line represents a differential excition interval, and each row represents a principal direction, and each fritter represents an interval ξ of the differential excition determined _mwith principal direction Φ _t.For the ease of classification, 2D histogram is converted to one dimensional histograms, histogrammic every a line composition one dimensional histograms H (k) of 2D, (k=0,1 ..., M-1), the interval ξ of the corresponding differential excition of each sub-histogram H (k) _k, all sub-histograms couple together and obtain whole one dimensional histograms H={H _m, m=0,1 ..., M-1.

Step (4) in PCA method (Principal Components Analysis) dimensionality reduction be more owing to improving the local feature that weber local description extracts, intrinsic dimensionality is higher, so the object in order to realize dimensionality reduction, the major component in characteristics of image is chosen out.Choose a front p maximal eigenvector and characteristic of correspondence thereof according to the contribution rate of eigenwert, get a=99%.

If, the eigenmatrix x=(x of all images ₁, x ₂..., x _n) ^t

Xi is that each column vector of i-th image is piled into row and obtains, namely the eigenmatrix vectorization of every sub-picture;

First, the average image is calculated

Then, calculated difference image M _i=x _i-Ψ, i=1,2 ..., N

3rd step, builds and obtains covariance matrix S

S＝M×M ^T

Wherein, M=(M ₁, M ₂..., M _n);

Finally, svd is carried out to the covariance matrix of structure, obtain eigenwert and characteristic of correspondence vector, choose a front p maximal eigenvector and characteristic of correspondence vector thereof according to the contribution rate of eigenwert, the space be made up of proper vector is exactly the optimum projection matrix that PCA obtains.Last reduction process is exactly a projection process:

$y_k=W_{\mathrm{opt}}^T\&times;x_k,k=\mathrm{1,2},...,N$

The new feature vector yk of low-dimensional can be converted to for the eigenvector xk that each M × N ties up before.

The distance of step (5) in three rank nearest neighbour classifications is the concept the most intuitively of measurement two image similarity degree.Represent the distance between sample x and y with L (x, y), the dimension of feature is k, obtains distance function below.Minkowsky range formula is:

$L(x,y)={\left[\underset{i=1}{\overset{k}{\mathrm{\&Sigma;}}}{|x_i-y_i|}^{\mathrm{\&lambda;}}\right]}^{\frac{1}{\mathrm{\&lambda;}}}$

Adopt three rank nearest neighbour methods to calculate the distance in lower dimensional space between image, the distance of three rank nearest neighbour methods is manhatton distances, the Minkowsky distance namely when λ=1, and range formula is:

$L(x,y)=\underset{i=1}{\overset{k}{\mathrm{\&Sigma;}}}|x_i-y_i|$

Final recognition result is obtained by three rank nearest neighbour classifications.

Although disclose embodiments of the invention and accompanying drawing for the purpose of illustration, but it will be appreciated by those skilled in the art that: in the spirit and scope not departing from the present invention and claims, various replacement, change and amendment are all possible, therefore, scope of the present invention is not limited to the content disclosed in embodiment and accompanying drawing.

Claims (5)

1., based on the face identification method simplifying flow model and improvement weber local description, it is characterized in that: comprise the steps:

(1) thermogram conversion rheography: thermogram is converted to rheography by the flow model simplified, obtains the blood-stream image of face;

(2) Gaussian smoothing: carry out Gaussian smoothing operation to above-mentioned blood-stream image, removes the interference of noise, improves infrared face recognition performance;

(3) improve weber local description and extract local feature: by the weber local description improved, local shape factor is carried out to the rheography of piecemeal, wherein the calculating of directional information utilizes with tropism Sobel operator, obtain more stable local authentication information, the WLD be improved, (1) (3) completed the WLD of the WLD of the improvement in training image and the improvement of test pattern by step to step respectively, obtain feature database image information and test pattern information respectively;

(4) PCA method dimensionality reduction: respectively dimensionality reduction is carried out to the WLD that the WLD improved in training image and test pattern improve;

(5) identify: the distance in the lower dimensional space calculating the feature database image information after above-mentioned dimensionality reduction and test pattern information between image, will obtain final recognition result by three rank nearest neighbour classifications.

2. the face identification method based on simplifying flow model and improvement weber local description according to claim 1, is characterized in that: the flow model of the simplification that step is (1) described is as follows:

$\mathrm{\&omega;}=\frac{\mathrm{\&epsiv;\&sigma;}(T^4-{T_e}^4)}{\mathrm{\&alpha;}{c}_b(T_a-T)}$

Wherein, the implication of each physical descriptor is: T _afor artery temperature, T _efor environment temperature, T are skin temperature, ε, σ, α, c _bfor constant parameter.

By simplifying the conversion of flow model, obtain the rheography of face.

3. according to claim 1 based on simplification flow model and the face identification method improving weber local description, it is characterized in that: the Gaussian smoothing employing following formula calculating that step is (2) described:

I'＝I*G(x,y,σ)

Wherein, * represents convolution algorithm, and gaussian kernel function is as follows:

$G(x,y,\mathrm{\&sigma;})=\frac{1}{2\mathrm{\&pi;}{\mathrm{\&sigma;}}^2}\exp (-\frac{x^2+y^2}{2{\mathrm{\&sigma;}}^2})$

Wherein σ is the variance of Gaussian filter.

4. the face identification method based on simplifying flow model and improvement weber local description according to claim 1, it is characterized in that: the (3) described improvement weber local description of step extracts local feature and is: first carry out piecemeal to facial image, face is divided into i × j sub-image area; Then the weber local description improved is utilized to extract local feature to each subimage respectively, the pixel differential excition value of computed image, by the differential excition value sum of surrounding pixel and center pixel than upper center pixel value, obtain a ratio, therefrom can obtain the marked change information of topography, calculate surrounding pixel and the difference sum of center pixel and the ratio of center pixel and adopt following formula:

$G_{\mathrm{ratio}}\left(x_c\right)=\frac{v_s^{00}}{v_s^{01}}$

Wherein v ₀₀and v ₀₁differential excition operator f respectively ₀₀and f ₀₁output, v ₀₀represent the differential excition value sum of surrounding pixel and center pixel, v ₀₁represent original image, x _cit is center pixel.

Do the differential excition that arctan conversion obtains Current central pixel again:

$\mathrm{\&xi;}\left(x_c\right)=\arctan \left[\frac{v_s^{00}}{v_s^{01}}\right]=\arctan \left[\underset{i=0}{\overset{p-1}{\mathrm{\&Sigma;}}}\left(\frac{x_i-x_c}{x_c}\right)\right]$

This differential excition ξ (x _c) variation range exist in order to form the histogram that can add up, M interval is quantized in differential excition, ξ _m(m=0,1 ... M-1).For each interval ξ _m, have ξ _m=[d _m, u _m], wherein lower interval d _m=(m/M-1/2) π, upper interval u _m=[(m+1)/M-1/2] π.

Directional information described by direction, the graded ratio of image level direction and vertical direction, introduce isotropy Sobel operator and replace original method to carry out compute gradient directional information, the computing method of the gradient direction of isotropy Sobel operator are as follows:

$\mathrm{\&theta;}\left(x_c\right)=\arctan \left(\frac{f_s^{11}}{f_s^{10}}\right)$

$f_s^{10}=f_{10}*I,$ $f_s^{11}=f_{11}*I$

Direction θ (x _c) variation range be the same with differential excition, in order to better set up histogram, T principal direction is quantized in the change in direction.Do before a quantization and map f as follows:

${\mathrm{\&theta;}}^{\&prime;}=\arctan 2(v_s^{11},v_s^{10})+\mathrm{\&pi;},$

$\arctan 2(v_s^{11},v_s^{10})=\left\{\begin{array}{cccc}\mathrm{\&theta;},& v_s^{11}>0& \mathrm{and}& v_s^{10}>0,\\ \mathrm{\&pi;}+\mathrm{\&theta;},& v_s^{11}>0& \mathrm{and}& v_s^{10}<0,\\ \mathrm{\&theta;}-\mathrm{\&pi;},& v_s^{11}<0& \mathrm{and}& v_s^{10}<0,\\ \mathrm{\&theta;},& v_s^{11}<0& \mathrm{and}& v_s^{10}>0,\end{array}\right.$

The expanded range of direction change is to [0,2 π].Quantization function is as follows:

After obtaining differential excition and direction, build the histogram of a 2D, { WLD (ξ _m, Φ _t); M=0,1 ..., M-1; T=0,1 ..., T-1; M and T represents the number in differential excition interval and the number of principal direction respectively, the 2D histogram obtained, and every a line represents a differential excition interval, and each row represents a principal direction, and each fritter represents an interval ξ of the differential excition determined _mwith principal direction Φ _t.For the ease of classification, 2D histogram is converted to one dimensional histograms, histogrammic every a line composition one dimensional histograms H (k) of 2D, (k=0,1 ..., M-1), the interval ξ of the corresponding differential excition of each sub-histogram H (k) _k, all sub-histograms couple together and obtain whole one dimensional histograms H={H _m, m=0,1 ..., M-1.

5. the face identification method based on simplifying flow model and improvement weber local description according to claim 1, it is characterized in that: the (5) described identification of step is: with L (x, y) represent the distance between sample x and y, the dimension of feature is k, obtains distance function below.Minkowsky range formula is:

$L(x,y)={\left[\underset{i=1}{\overset{k}{\mathrm{\&Sigma;}}}{|x_i-y_i|}^{\mathrm{\&lambda;}}\right]}^{\frac{1}{\mathrm{\&lambda;}}}$

Adopt three rank nearest neighbour methods to calculate the distance in lower dimensional space between image, the distance of three rank nearest neighbour methods is manhatton distances, the Minkowsky distance namely when λ=1, and range formula is:

$L(x,y)=\underset{i=1}{\overset{k}{\mathrm{\&Sigma;}}}|x_i-y_i|$

Final recognition result is obtained by three rank nearest neighbour classifications.