CN205788213U - A kind of face identification device - Google Patents

Abstract

The utility model discloses a kind of face identification device, described device includes: image collection module, sparse coding module, joint classification module and label output module, above-mentioned each module is linked in sequence successively, wherein, described image collection module, the face image data being used for acquiring is as test sample；Described sparse coding module, for described test sample is carried out sparse coding, obtains the sparse expression of test sample；Described joint classification module, including dictionary grader and linear classifier, for described test sample is carried out joint classification；Described label output module, for exporting the class label of described test sample.Test sample is classified by above-mentioned face identification device, associating dictionary grader and linear classifier, it is possible to better profit from the discriminant information that sparse vector comprises, and strengthens classification capacity, improves recognition success rate.

Description

A kind of face identification device

Technical field

This utility model relates to computer vision and area of pattern recognition, particularly relates to a kind of face identification device.

Background technology

In recent years due to the proposition of rarefaction representation grader (Sparse Representation-based classification is called for short SRC) method, the research of recognition of face achieves the progress attracted attention.

The task of recognition of face can be such defined that a given facial image (test sample), it is judged which individual in the corresponding image (training sample set, also referred to as dictionary) collected in advance of the identity of this facial image.Wherein comprising the people of multiple different identity inside training sample set, everyone has multiple training samples.The foundation of SRC method is: to each test sample, and it can be represented by the linear combination of the facial image of training sample concentration identical category.Like this, the coefficient that the training sample identical with test sample classification is corresponding is not zero, and coefficient corresponding to other sample is zero.Therefore, this sparse expression vector, the test sample a kind of expression in training set, also referred to as rarefaction representation, and this expression can be regarded as and reflected the identity information of test image well.

SRC method may be summarized to be two steps: (1) solves the rarefaction representation of corresponding test sample by the existing solution of convex optimization；(2) according to the rarefaction representation obtained, dictionary grader (Dictionary Classifier is called for short DC) classification is utilized.A large amount of theoretical researches show with experiment, and SRC method is more superior than traditional face identification method (such as nearest neighbor method, vector machine method etc.) performance.At present relevant SRC technique study be concentrated mainly on train the study of dictionary, sparse algorithm for reconstructing, to blocking, mismatch is accurate and the aspect such as single sample scene.

The difficult point of recognition of face at present is:

(1) the face plastic deformation that expression causes

(2) the face multiformity that attitude causes

(3) the face change that the age causes

(4) multiplicity of the face pattern that the factor such as hair style, beard, glasses, cosmetic causes

(5) factor that the diversity of the facial image that the factor such as the angle of illumination, intensity and sensor characteristics causes is many makes face identification rate significantly decline.

Although the DC in SRC method has been achieved for preferable experiment effect, but due to recognition of face problem complexity in reality scene, DC yet suffers from a lot of limitation, causes the failure of identification mission.

Summary of the invention

The purpose of this utility model is to provide the method and device of a kind of recognition of face, it is therefore intended that strengthen existing identification technology classification ability, improves recognition success rate, preferably carries out recognition of face.

For solving above-mentioned technical problem, this utility model provides a kind of face identification device, and described device includes: image collection module, sparse coding module, joint classification module and label output module, and above-mentioned each module is linked in sequence successively, wherein,

Described image collection module, the face image data being used for acquiring is as test sample；

Described sparse coding module, for described test sample is carried out sparse coding, obtains the sparse expression of test sample；

Described joint classification module, including dictionary grader and linear classifier, for described test sample is carried out joint classification；

Described label output module, for exporting the class label of described test sample.

Further, described device also includes that training module, described training module are connected with described joint classification module, constructs described dictionary grader and the described linear classifier of training in advance.

Further, described training module includes: training sample set sets up unit, dictionary grader structural unit and linear classifier training unit, and above-mentioned each unit is linked in sequence successively, wherein,

Described training sample set sets up unit, is used for setting up face training sample set Y；

Described dictionary grader structural unit, is used for utilizing described face training sample set Y to construct complete dictionary D；

Described linear classifier training unit, is used for utilizing described face training sample set Y training linear classifier W.

Further, described joint classification module includes: dictionary grader, the first judgement unit, linear classifier and the second judgement unit, and above-mentioned each unit is linked in sequence successively, wherein,

Described dictionary grader, for carrying out dictionary classification to test sample；

Described first judgement unit, for the first residual index RI by structure₁Judge that the classification results of dictionary grader is the most reliable；

Described linear classifier, for carrying out linear classification to test sample；

Described second judgement unit, for the second residual index RI by structure₂Judge that the classification results of linear classifier is the most reliable.

Further, described first residual index RI₁Can according to test sample be under the jurisdiction of target sample posterior probability structure obtain, specific as follows:

First residual index RI of definition dictionary grader₁:

${\mathrm{RI}}_1\left(\stackrel{\→}{y_i}\right|y)=\frac{1}{{\mathrm{\μ}}_1}\exp \{-{\mathrm{\μ}}_2\·||\stackrel{\→}{y_i}-\stackrel{\→}{y}|{|}_2^2\}$

Wherein,For to RI₁It is normalized, μ₂It it is constant.

Further, described second residual index RI₂Can according to test sample be under the jurisdiction of target sample posterior probability structure obtain, specific as follows:

Second residual index RI of definition linear classifier₂:

${\mathrm{RI}}_2\left(\stackrel{\→}{z_i}\right|z)=\frac{1}{{\mathrm{\μ}}_1}\exp \{-{\mathrm{\μ}}_2\·||\stackrel{\→}{z_i}-\stackrel{\→}{z}|{|}_2^2\}$

Wherein,For to RI₂It is normalized, μ₂It it is constant.

The present invention has such advantages as relative to prior art and effect:

The face identification device that this utility model is proposed, carries out sparse coding to face test sample, obtains the sparse vector of test sample.Then utilize dictionary grader and linear classifier successively test sample to be classified, determine the classification of test sample according to kind judging formula, to reach the purpose of recognition of face.Test sample is classified by face identification device provided by the utility model, associating dictionary grader and linear classifier, it is possible to better profit from the discriminant information that sparse vector comprises, and strengthens classification capacity, improves recognition success rate.

Accompanying drawing explanation

In order to be illustrated more clearly that the technical scheme of this utility model embodiment, in describing embodiment below, the required accompanying drawing used is briefly described.

Fig. 1 is the structured flowchart of the face identification device disclosed in this utility model；

Fig. 2 is another structured flowchart of the face identification device disclosed in this utility model；

Fig. 3 is the structured flowchart of training module in the face identification device disclosed in this utility model；

Fig. 4 is the structured flowchart of joint classification module in the face identification device disclosed in this utility model；

Fig. 5 is that the nicety of grading of three-type-person's face identification device is along with atomic number change curve.

Detailed description of the invention

Below in conjunction with the accompanying drawing in this utility model embodiment, the technical scheme in this utility model embodiment is clearly and completely described, it is clear that described embodiment is only a part of embodiment of this utility model rather than whole embodiments.Based on the embodiment in this utility model, the every other embodiment that those of ordinary skill in the art are obtained under not making creative work premise, broadly fall into the scope of this utility model protection.

Embodiment

The structured flowchart of a kind of detailed description of the invention of face identification device provided by the utility model is as it is shown in figure 1, this device includes:

Image collection module 1, the face image data being used for acquiring is as test sample.

Sparse coding module 2, for described test sample is carried out sparse coding, obtains the sparse expression of test sample.

Joint classification module 3, including dictionary grader and linear classifier, for described test sample being carried out joint classification, to improve recognition success rate.

Label output module 4, for exporting the class label of described test sample.

In another preferred embodiment, the structured flowchart of the another kind of detailed description of the invention of face identification device provided by the utility model is as shown in Figure 2, this device also includes: training module 5, dictionary grader and training linear classifier is constructed in advance, described test sample is classified, to improve recognition success rate.

Training module 5 in the device of recognition of face provided by the utility model may include that further

Training sample set sets up unit 501, is used for setting up face training sample set Y.

Dictionary grader structural unit 502, is used for utilizing described face training sample set to construct complete dictionary D.

Linear classifier training unit 503, is used for utilizing described face training sample set training linear classifier W.The structured flowchart of above-mentioned training module 5 is as shown in Figure 3.Wherein, the work process of linear classifier training unit 503 is:

The object function of definition linear classifier is:

W=argmin_W,X||Y-DX||_F+λ₁||X||₁+λ₂||H-WX||_F+λ₃||W||_F

Wherein,It is training image collection,Being the rarefaction representation coefficient matrix of training image collection, D was complete dictionary, column vector h of H_i=[0,0 ..., 1 ..., 0,0]^T∈R^{k × 1}Being the label vector of training image, position element corresponding to each classification is not zero, | | Y-DX | |_FIt is to represent error, | | H-WX | |_FIt is error in classification, | | X | |₁It is sparse constraint item, | | W | |_FIt is that item, λ are penalized in regularization₁, λ₂, λ₃It it is the scalar maintaining every balance.

Algorithm flow is:

(1) Y, D, H are initialized；

(2) ignore Section 3 and the Section 4 of object function, solved by following formula and represent coefficient matrix X:

W=| | Y-DX | |_F+λ₁||X||₁

(3) ignore Section 1 and the Section 2 of object function, utilize and solve the expression coefficient matrix X obtained, solve linear classifier W by following formula:

W=λ₂||H-WX||_F+λ₃||W||_F

(4) output X, W.

In another preferred embodiment, as shown in Figure 4, this joint classification module 3 farther includes the structured flowchart of the joint classification module 3 in face identification device provided by the utility model:

Dictionary grader 301, for test sample carries out dictionary classification, and exports classification results and carries out reliability judgement to the first judgement unit 302；

First judgement unit 302, for the first residual index RI by structure₁((Residual Index, be called for short RI) judges the classification results of dictionary grader the most reliably, if classification results is judged as reliably, then going to label output module 4, if classification results is judged as unreliable, then goes to linear classifier 303；

Linear classifier 303, for test sample carries out linear classification, and exports classification results and carries out reliability judgement to the second judgement unit 304；

Second judgement unit 304, for the second residual index RI by structure₂Judge that the classification results of linear classifier is the most reliable, if classification results is judged as reliably, then going to label output module 4.

Further, residual index RI can be specially according to test sample be under the jurisdiction of target sample posterior probability structure obtain:

First residual index RI of definition dictionary grader₁:

${\mathrm{RI}}_1\left(\stackrel{\→}{y_i}\right|y)=\frac{1}{{\mathrm{\μ}}_1}\exp \{-{\mathrm{\μ}}_2\·||\stackrel{\→}{y_i}-\stackrel{\→}{y}|{|}_2^2\}$

Wherein,For to RI₁It is normalized, μ₂It it is constant；

Second residual index RI of definition linear classifier₂:

${\mathrm{RI}}_2\left(\stackrel{\→}{z_i}\right|z)=\frac{1}{{\mathrm{\μ}}_1}\exp \{-{\mathrm{\μ}}_2\·||\stackrel{\→}{z_i}-\stackrel{\→}{z}|{|}_2^2\}$

Wherein,For to RI₂It is normalized, μ₂It it is constant.

Introducing the engineering process of the face identification device that this utility model provides in detail below, in this implementing procedure, Extend Yale B face database contains 2432 facial images of 38 people.Every image shoots under different illumination conditions, and the size of image is 195 × 168 pixels.Randomly choosing 50% as training sample from data base, remaining 50%, as test sample, is down-sampled to 120 dimensions to each image, and the atomic number of each class of dictionary is taken as 4,8,12,16 respectively, carries out 5 tests and averages.

Specifically, the present embodiment includes constructing face training dataset, structure dictionary grader and training linear classifier and the process utilizing both the above grader to classify image, comprises the following steps:

Set up face training sample set Y ∈ R^{m × n}；

According to face training sample set Y, constructed complete dictionary D；

According to face training sample set Y, training linear classifier W；

Input face test sample y；

To crossing complete dictionary D and test sample y, utilize following majorized function:

${\hat{x}}_1=\arg {\min }_x\left|\right|x|{|}_{1}s.t.Dx=y$

Solve the rarefaction representation coefficient x of optimum；

The rarefaction representation coefficient x obtained will be solved, substitute into classification band pass function δ respectively_i, test sample is reconstructed, the sample after being reconstructedWherein i=1,2 ..., k, representative sample classification；

Calculate the class reconstructed error r of the sample after reconstruct and test sample_i:

$r_i=\left|\right|y-{\hat{y}}_i|{|}_2$

By the class reconstructed error r of test sample_iSubstitution kind judging formula:

Identity (y)=argmin_i r_i

Try to achieve the minima in k residual error, and using its subscript i as the recognition result of dictionary grader, represent with identity (y)；

Residual index RI of Dictionary of Computing classification^DCIf, RI^DC(y) > τ, output identity (y), as recognition result, otherwise, it determines dictionary classification results is unreliable, continues with linear classifier and classifies；

The linear classifier W obtained, employing following formula calculating test sample y is relative to the Weighted Similarity matrix of each classification:

$Z=W{\hat{x}}_1={\[z_1,z_2,...,z_k\]}^T$

The Weighted Similarity substitution kind judging formula of test sample y:

Identity (y)=argmax_iz_i

Try to achieve the maximum in k Weighted Similarity, and using its subscript i as the recognition result of linear classifier, represent with identity (y)；

Calculate residual index RI of linear classification^LC, threshold θ ∈ (0,1) is set, if RI^LC(y) > θ, output identity (y) is as recognition result.

Export the class label of facial image to be identified.

This completes the classification to facial image to be identified.

Fig. 5 gives the nicety of grading of three kinds of algorithms along with dimension change curve.Three kinds of control methods are respectively as follows: SRC, Discriminative K-SVD (being called for short DKSVD) and this utility model.It can be seen that in the case of atom number is fewer, discrimination of the present utility model, apparently higher than other two kinds of methods, therefore can bring convenience realizing when.

Described above to the disclosed embodiments, makes professional and technical personnel in the field be capable of or uses this utility model.Multiple amendment to embodiment will be apparent from for those skilled in the art, and generic principles defined herein can realize in the case of without departing from spirit or scope of the present utility model in other embodiments.Therefore, this utility model is not intended to be limited to embodiment illustrated herein, and is intended to accord with principles disclosed herein and the consistent the widest scope of features of novelty.

Claims (5)

1. a face identification device, it is characterised in that described device includes: image collection module, sparse coding module, joint classification module and label output module, above-mentioned each module is linked in sequence successively, wherein,

Described image collection module, the face image data being used for acquiring is as test sample；

Described sparse coding module, for described test sample is carried out sparse coding, obtains the sparse expression of test sample；

Described joint classification module, including dictionary grader and linear classifier, for described test sample is carried out joint classification；

Described label output module, for exporting the class label of described test sample；

Described device also includes that training module, described training module are connected with described joint classification module, constructs described dictionary grader and the described linear classifier of training in advance.

A kind of face identification device the most according to claim 1, it is characterised in that described training module includes: training sample set sets up unit, dictionary grader structural unit and linear classifier training unit, and above-mentioned each unit is linked in sequence successively, wherein,

Described training sample set sets up unit, is used for setting up face training sample set Y；

Described dictionary grader structural unit, is used for utilizing described face training sample set Y to construct complete dictionary D；

Described linear classifier training unit, is used for utilizing described face training sample set Y training linear classifier W.

A kind of face identification device the most according to claim 1, it is characterised in that described joint classification module includes: dictionary grader, the first judgement unit, linear classifier and the second judgement unit, and above-mentioned each unit is linked in sequence successively, wherein,

Described dictionary grader, for carrying out dictionary classification to test sample；

Described first judgement unit, for the first residual index RI by structure₁Judge that the classification results of dictionary grader is the most reliable；

Described linear classifier, for carrying out linear classification to test sample；

Described second judgement unit, for the second residual index RI by structure₂Judge that the classification results of linear classifier is the most reliable.

A kind of face identification device the most according to claim 3, it is characterised in that described first residual index RI₁The posterior probability structure being under the jurisdiction of target sample according to test sample obtains, specific as follows:

First residual index RI of definition dictionary grader₁:

Wherein,For to RI₁It is normalized, μ₂It it is constant.

A kind of face identification device the most according to claim 3, it is characterised in that described second residual index RI₂The posterior probability structure being under the jurisdiction of target sample according to test sample obtains, specific as follows:

Second residual index RI of definition linear classifier₂:

Wherein,For to RI₂It is normalized, μ₂It it is constant.