CN105893916A - New method for detection of face pretreatment, feature extraction and dimensionality reduction description - Google Patents

Abstract

The present invention provides a new method for detection of face pretreatment, feature extraction and dimensionality reduction description. The problems of face correction, face extraction and face feature dimension reduction at face identification are solved. The new method is able to effectively improve the identification stability of a feature extraction algorithm by using face images collected in the natural environment through a series of pretreatment process. Especially, the new method is able to greatly solve the great reduction problem of the identification precision caused by face attitude changing, illumination changing and fuzzy characteristics, and has application values in practice. Through combination of the subspace learning mode, the new method is able to effectively perform compact compression of the extracted descriptors through combination of the Gabor and the LPQ in the expression process from high-dimensional face feature to low-dimensional subspace mapping so as to improve the face comparison calculation efficiency.

Description

A kind of new method that the detection pretreatment of face, feature extraction to dimensionality reduction are described

Technical field

Patent of the present invention relates to CRT technology and intelligence system technical field, especially a kind of The new method that the detection pretreatment of face, feature extraction to dimensionality reduction are described.

Background technology

Biometrics identification technology, such as recognition of face, is to utilize the facial image got to carry out body The application technology that part identifies.So-called biological characteristic, refers to the built-in attribute existing for face, tool There is the otherness between stronger stability and individuality.Typical living creature characteristic recognition system is to utilize The facial image obtained is compared with the image in database and exports comparison result.

The object of recognition of face generally comprises various certificate photo.Such as identity card, student's identity card, passport with And the natural image of photographing outdoors, in numerous facial images, the particularly people under natural environment Face image, it has more attitudes vibration, such as eye position not at same level line, and, Information effective to recognition of face is generally focused on blee region, hair and neck area pair The effect that precision improves is little, on the contrary, may affect accuracy of identification.Therefore, carry at face characteristic Before taking the stage, it usually needs be corrected facial image, this is that recognition of face is urgently to be resolved hurrily Key issue.

During the analysis to facial image, the otherness obtained due to vision facilities and image The difference of quality, adds some natural trend that face self exists, such as expression, attitude, illumination The impact at change, age etc..The impact brought facial image for illumination variation, generally to rear Continuous accuracy of identification has bigger challenge.As illumination effect is the most serious, recognition of face accurate The degree more end, in recent years, owing to recognition of face gradually walks out laboratory environment and toward oneself under natural environment Dynamic recognition of face development, and under natural environment, illumination effect is a very important factor, Therefore, the facial feature extraction of illumination invariant is recognition of face problem demanding prompt solution.

Although face recognition technology has been achieved for good achievement, but current most recognitions of face Technology designs and in terms of model training the most often just for the situation that picture quality is good at algorithm, and right For intelligent monitoring, public security system suspect's image are than peer application, due to facial image Differing in source, the quality of some image is excessively poor, such as fuzzy, strong noise, resolution ratio are low, Both increase the difficulty of image recognition, how to improve the system facial characteristics to these low-quality images Identify that extractability is also one of recognition of face key issue urgently to be resolved hurrily.

When the mode using Gabor filtering extracts the feature of facial image, the dimension of feature is usual The highest, the actual application of face alignment identification also has higher computation complexity, It is unfavorable for large-scale facial image identification and analysis.The dimension how reducing feature is also that face is known Key issue not urgently to be resolved hurrily.

Patent of invention content

Extract to solve face normalization, face characteristic when recognition of face and reduce face characteristic dimension The problem of degree, the present invention provides a kind of and describes the detection pretreatment of face, feature extraction to dimensionality reduction New method.

The technical solution adopted for the present invention to solve the technical problems is:

In conjunction with the several key technical problems related in Automatic face recognition system, propose one in detail Thin handling process, after detection face input, positions (mainly eyes to face feature point Location)；According to the eye position navigated to, fixed proportion is used to cut out the people of normalization size Face image, to normalized facial image, carries out illumination pretreatment, Gamma correction, rejects The illumination effect that imaging device or natural environment are brought, improves stablizing of follow-up face characteristic extraction Property.In conjunction with Gabor filtering and the mode of LPQ partial descriptions, the image of illumination pretreatment is carried The local description of consistency stuck with paste by delivery, effectively describes the change of face texture feature.Finally, knot Zygote space learning, the high dimensional feature that Gabor Yu LPQ is extracted, use speced learning Mode, is divided into multistage by high dimensional feature, respectively each section of Feature Descriptor is carried out sub-space learning And obtain the feature after dimensionality reduction, finally the feature after multistage dimensionality reduction is formed complete face and describe Son, and use the mode of cosine similarity measurement to carry out comparison and the identification of face.

Patent of the present invention provides the benefit that:

1, the facial image collected under natural environment, by a series of preprocessing process, It is effectively improved the identification stability of feature extraction algorithm.Particularly in, illumination changeable to human face posture The accuracy of identification that changeable, fuzzy behaviour is brought significantly declines problem, has good improvement result, The most more using value.

2, the mode of zygote space learning, maps from higher-dimension face characteristic to lower-dimensional subspace Expression process, description that can effectively combine extracted to Gabor with LPQ carries out compact Compression, improves the computational efficiency of face alignment.

Accompanying drawing explanation

Fig. 1 is a kind of new method stream describing the detection pretreatment of face, feature extraction to dimensionality reduction Cheng Tu

Detailed description of the invention

Process is face normalization, facial pretreatment, face characteristic extraction, the low dimension table of high dimensional feature Showing, details is as follows:

1, face normalization

Design multiple adaboost¹Human eye detection grader.Facial image is carried out eye detection it Before, if picture size is less, adaboost testing result may be caused undesirable, thus generally need Picture size is carried out interpolation amplification to improve resolution ratio, be so favorably improved adaboost classification The locating effect of device.Utilize multiple eye detection grader, image is carried out eye detection, training The acquisition of file, can be by opencv²The grader of the thread provided, it is possible to according to particular demands, Collect training sample voluntarily and train grader.Detailed process is as follows: 1. (logical according to multiple graders More satisfactory frequently with 3～6 meetings) testing result, the alternative point of multiple human eye can be obtained.2. limit Determining the eye position bound at the longitudinal axis, alternative in bound is clicked on by recycling K-mean algorithm The cluster process of row two class, two final cluster centres represent images of left and right eyes position respectively.3. will The perpendicular bisector of the line of two cluster centres is as the cut-off rule of images of left and right eyes, to two clusters Center, finds the alternative point nearest with each center, respectively as final positioning result.Human eye Detection completes.

2, facial pretreatment

According to the eyes position navigated to, fixed proportion is used to cut out standardized face and carry out Illumination pretreatment, pretreatment process comprises 3 step: Gamma corrections, DOG (difference Of gauss) filter and to comparison normalization operation.Gamma correction be one nonlinear Image gray-scale transformation, for image I, generally uses nonlinear transformation such as I^γ(γ ＞ 0) or log (I), Each pixel grey scale in image is converted.Gamma operation enhances relatively dark portion in image The gray scale dynamic range divided, has also carried out certain compression to high bright part simultaneously.At Gamma On the basis of correction, by DOG filtering operation, can effectively extract the HFS in image, Further, the edge contour feature of facial image is further enhanced.On the basis of single Gaussian kernel, By using two different σ that image carries out the convolutional filtering of different scale, different by two width The image difference of scale filter result, the difference image of available DOG:

$G_{\mathrm{\σ}}(x,y)=\frac{1}{\sqrt{2{\mathrm{\π\σ}}^2}}\exp [-\frac{x^2+y^2}{2{\mathrm{\σ}}^{\text{2}}}]$

$I^{\′}=G_{{\mathrm{\σ}}_1}*I-G_{{\mathrm{\σ}}_2}*I=(G_{{\mathrm{\σ}}_1}-G_{{\mathrm{\σ}}_2})*I=\mathrm{DoG}*I$

Last step of facial pretreatment is normalization operation.Normalization process is mainly passed through Below three processing stage, with by image normalization to (-τ, τ) scope:

$I(x,y)=\frac{I(x,y)}{{\left(\mathrm{avg}\left({\left|I\right|}^{\mathrm{\α}}\right)\right)}^{1/\mathrm{\α}}}$

$I(x,y)=\frac{I(x,y)}{{\left(\mathrm{avg}\left(\min (\mathrm{\τ},{\left|I\right|}^{\mathrm{\α}})\right)\right)}^{1/\mathrm{\α}}}$

$I(x,y)=\mathrm{\τ}\tan \left(\frac{I(x,y)}{\mathrm{\τ}}\right)$

Wherein, α is a stronger compressibility factor, very big with produce after reducing DOG filtering Value.τ is a threshold value, for blocking maximum.Entire image is taken all by avg () expression Value.Min () represents the minimum of a value of numerical value and image pixel.

3, face characteristic extracts

Gabor³Small echo can simple mononeuric receptive field type be also in analog vision cortex well Extract significant visual characteristic, such as space orientation, set direction etc..Particularly Gabor wavelet Spatial frequency multiple dimensioned, multi-direction extract feature so that can divide from the angle of overall situation and partial situation Analysis face characteristic.To the facial image after illumination pretreatment, use the mode of Gabor filtering, carry Take multiple dimensioned, the Orientation Features of face.Have chosen 5 yardsticks, the wave filter in 8 directions. On the amplitude response of Gabor filtering output, extract LPQ feature, to obtain on local grain Fuzzy consistency describe.The principle of LPQ is: blurred picture (G) is by original image (F) Caused by the convolution of some diffusion (PSF) function (H), in frequency domain, show as Represent the dot product operation of frequency domain, further, if only considering phase spectrum, then ∠ G=∠ F+ ∠ H. If PSF is Central Symmetry, then ∠ H is 0 or π, ∠ H be 0 can make ∠ F have fuzzy not Becoming attribute, this fuzzy convolution process is the essence of LPQ.

In actual application, implement step as follows: face through illumination pretreatment, 5 After yardstick, 8 trend pass filterings, form 40 Gabor amplitude figures.Each width Gabor figure is carried out The division of 8*8 block, to ensure that the space structure of facial image is distributed.Each local Gabor is schemed As block extracts LPQ, son is described.Original LPQ describes sub by 8 binary representations, and it describes Scope is 0～255 (256 rank).The LQP of all localized masses is described son and uses series system. Due to the dimension disaster problem of feature, original 256 rank of LPQ can be carried out amount in various degree Change.Use the mode of uniform quantization, such as 8,16,32,64,128,256 rank.For above-mentioned various amounts Change mode, selects one.The LPQ feature interpretation extracted in conjunction with Gabor, to illumination, mould The face of paste property, has stronger robustness.

4, the low-dimensional of high dimensional feature represents

The subspace method for expressing of two kinds is provided, albefaction PCA (Whitening PCA), LDA, Select one.

For the subspace training process of single sample, on the basis using PCA to reduce data dimension On.Also needing to an associated pre-treatment step, this preprocessing process is referred to as albefaction.Pin For the LPQ local description of higher-dimension, owing to describing, there is between sub-dimension the strongest being correlated with Property, for training time input be redundancy.The purpose of Whitening reduces input Redundancy, is a lot of algorithm step of carrying out pre-processing.Learning algorithm is made by whitening process Input has the property that between (i) feature, correlation is relatively low；(ii) all features have identical side Difference.Common whitening operation has PCA whitening and ZCA whitening.Use PCA Whitening, refers to data x after PCA dimensionality reduction is z, it can be seen that the most one-dimensional in z be Independent, meet first condition of whitening albefaction, this is have only in z each It is 1 that dimension has all just obtained every one-dimensional variance divided by standard deviation, say, that variance is equal.Formula For:

$x_{\mathrm{PCAwhite},i}=\frac{x_{\mathrm{rot},i}}{\sqrt{{\mathrm{\λ}}_i}}$

In actual list specimen discerning is applied, the dimensionality reduction accessed by albefaction PCA reduction process is special Levy, when using the metric form of cosine similarity (cosine similarity) during identifying, relatively Classical PCA dimensionality reduction identification, has more significantly accuracy benefits.Cosine similarity is to pass through The cosine of an angle value measured between two vectors measures the similitude between them.0 degree of angle remaining String value is 1, and the cosine value of other any angles is all not more than 1；And its minimum of a value is-1.From And the cosine value of the angle between two vectors determines whether two vectors are pointed generally in identical side To.Assuming that A and B is two n-dimensional vectors, A be [A1, A2 ..., An], B be [B1, B2 ..., Bn], then the cosine similarity of the angle theta of A with B is equal to:

$\cos \mathrm{\θ}=\frac{{\mathrm{\Σ}}_{i=1}^n(A_i\×B_i)}{\sqrt{{\mathrm{\Σ}}_{i=1}^n{\left(A_i\right)}^2}\×\sqrt{{\mathrm{\Σ}}_{i=1}^n{\left(B_i\right)}^2}}=\frac{A\·B}{\left|A\right|\×\left|B\right|}$

For multisample subspace train process, use LDA mode, study class in and class Between relation, LAD is different from PCA, LDA so that the separability of sample is preferably target, seeks Look for one group of optimum linear transformation to make within-cluster variance minimum, and inter _ class relationship is maximum, theoretical Upper theory, it is adaptable to pattern recognition problem.Classical LDA uses Fisher criterion function, its It is defined as follows:

$J\left(w\right)=\underset{w}{\arg \max }\frac{\left|w^T{S}_{B}w\right|}{\left|w^T{S}_{w}w\right|}$

Wherein, inter _ class relationship S_bWith S in class_wDispersion may be defined as:

$S_B=\frac{1}{C(C-1)}\underset{i=1}{\overset{C}{\mathrm{\Σ}}}\underset{j=1}{\overset{C}{\mathrm{\Σ}}}({\mathrm{\μ}}_i-{\mathrm{\μ}}_j){({\mathrm{\μ}}_i-{\mathrm{\μ}}_j)}^T$

$S_W=\frac{1}{C}\underset{i=1}{\overset{C}{\mathrm{\Σ}}}\frac{1}{N_i}\underset{j=1}{\overset{N_i}{\mathrm{\Σ}}}(x_j^i-{\mathrm{\μ}}_i){(x_j^i-{\mathrm{\μ}}_i)}^T$

C is total class number, N_iIt is expressed as the sample number of the i-th class,It is the jth sample in the i-th class,It it is the average of sample in the i-th class.Mathematically, Fisher criterion function is solved Optimal solution is equal to solveS_BEigenvalue problem.

For combining the LPQ Feature Descriptor that Gabor extracts, the reduction process of LDA and albefaction PCA is similar, again by the mode of segmentation, higher-dimension LPQ is described son and carries out multistage division, And each section of description is learnt its subspace respectively and carry out LDA dimensionality reduction, finally, by multistage LDA dimensionality reduction feature one characteristic vector of composition reaches to represent the dimensionality reduction of facial image.At face ratio Recognition of face is carried out to cognitive phase, the equally mode of employing cosine similarity.

The content of above detailed description of the invention is only the preferred embodiment of patent of the present invention, above-mentioned preferably Embodiment is not used for limiting the practical range of patent of the present invention；Every according to its power of patent of the present invention The various equivalents that the protection domain that profit requires is made, all by its claim of patent of the present invention Scope covered.

Claims (5)

1. the new method described the detection pretreatment of face, feature extraction to dimensionality reduction, it is special Levy and be: its process is face normalization, facial pretreatment, face characteristic extraction, high dimensional feature Low-dimensional represents.

The most according to claim 1 a kind of to detecting the pretreatment of face, feature extraction to fall The new method that dimension describes, it is characterised in that: face normalization needs to design the inspection of multiple adaboost human eye Survey grader.And picture size is carried out interpolation amplification to improve resolution ratio.Process is: 1. basis The testing result of multiple graders (generally 3～6 meetings of employing are more satisfactory), can obtain the most individual The alternative point of eye.2. limiting the eye position bound at the longitudinal axis, recycling K-mean algorithm is to up and down Alternative point in limit carries out the cluster process of two classes, two final cluster centres represent respectively a left side, Right eye position.3. using the perpendicular bisector of the line of two cluster centres as the segmentation of images of left and right eyes Line, to two cluster centres, finds the alternative point nearest with each center, respectively as final Positioning result.

The most according to claim 1 a kind of the detection pretreatment of face, feature extraction are arrived The new method that dimensionality reduction describes, it is characterised in that: facial pretreatment comprises 3 step: Gamma Correction, DOG (difference of gauss) filter and to comparison normalization operation.

The most according to claim 1 a kind of the detection pretreatment of face, feature extraction are arrived The new method that dimensionality reduction describes, it is characterised in that: use the mode of Gabor filtering to carry out LPQ people Face feature extraction.

Implement step as follows: face is through illumination pretreatment, 5 yardsticks, 8 trend pass filterings After, form 40 Gabor amplitude figures.Each width Gabor figure carries out the division of 8*8 block, with Ensure the space structure distribution of facial image.Each local Gabor image block is extracted LPQ retouch State son.Original LPQ describes sub by 8 binary representations, and it describes scope is 0～255 (256 Rank).The LQP of all localized masses is described son and uses series system.Dimension calamity due to feature Original 256 rank of LPQ can be carried out quantization in various degree by difficult problem.Use uniform quantization Mode, such as 8,16,32,64,128,256 rank.

The most according to claim 1 a kind of the detection pretreatment of face, feature extraction are arrived The new method that dimensionality reduction describes, it is characterised in that: the method using subspace, by higher-dimension face characteristic Carry out low-dimensional expression.

The subspace method for expressing of two kinds is provided, albefaction PCA (Whitening PCA), LDA, Select one.