Pre-detection and dual-dictionary sparse representation based face recognition algorithm in non-sufficient training samples
2018-03-07ZHAOJianZHANGChaoZHANGShunliLUTingtingSUWeiwenandJIAJian
ZHAO Jian,ZHANG Chao,ZHANG Shunli,LU Tingting,SU Weiwen,and JIA Jian
School of Information Science and Technology,Northwest University,Xi’an 710069,China
1.Introduction
Face recognition is a challenging and active research subject in computer vision,pattern recognition and artificial intelligence,which has been widely utilized in many fields such as image searching,personnel information security,and convicts detecting system[1–4].
In recent years,sparse representation classification(SRC)has been developed rapidly and played an important role in the field of face recognition.However,current algorithms based on the sparse representation also have some drawbacks.On one hand,test samples with various facial poses are one of the key obstacles of the face recognition algorithm.On the other hand,obtaining enough number of official facial images for some individual is not practical in daily life.In most cases,we just have fewer images than Scientific experiments database[5,6].In many contexts,such as finding suspects mission and collecting the evidence for traffic violation,the monitor may just gain two or three clear facial images.The performance is barely satisfactory under the condition of non-sufficient training samples,therefore the recognition rate is not satisfied.
In recent years,it has become a hot spot to explore the face recognition issue under the condition of non-sufficient training samples and facial poses.Many techniques have been used to solve the problem,including the neural network algorithms[7],synthesizing virtual samples[8],sparsity preserving discriminant analysis[9],and symmetrical face algorithm[10].
In our work,an algorithm is proposed to solve the problem mentioned above.The algorithm is able to improve the recognition rate under the condition of low training and various poses.Aiming at a higher recognition rate,we improve the robustness for the condition of various facial poses.The algorithm based on the traditional SRC is able to obtain better performance than the original algorithm under the low training samples condition.The basic frame of this paper is organized as follows.In Section 2,a pre-recognition algorithm based on geometry is proposed for various facial poses.In Section 3,a dual-dictionary algorithm based on the pre-recognition of facial poses and the framework of our recognition algorithm are introduced.In Section 4,we present experiments on olivetti research laboratory(ORL)and face recognition technology(FERET)databases.Section 5 gives the conclusion of our work.
2.Facial pose pre-recognition model
In this section,we present a novel approach that can im-prove the face recognition rate under the condition of complex and different poses.
In order to get better performance of the face recognition rate, different facial poses should be recognized and classified before constructing the dictionary. The pre-recognition algorithm for the facial pose is proposed based on the facial feature points and its geometry algorithm in the space domain.
2.1 Facial feature points detection
Pre-detection is one part of the facial pose pre-recognition(FPPR)model,whose purpose is to locate facial feature points.Various algorithms have been proposed for facial feature point detection,and this research field grows quickly.In recent years,the convolutional neural network(CNN)has become a novel way in facial feature detection since facial feature points can be rapidly and precisely detected by the CNN.Therefore,the CNN is utilized to perform the detection of facial feature points in pre-detection[11].
The core point of the FPPR algorithm is the geometry algorithm of facial feature points. Firstly, five facial feature points are obtained by the CNN.Then,using the geometry attributes of the five feature points, the facial pose classifier can be constructed.The five facial feature points are:left eye center,right eye center,nose tip,left mouth tip,right mouth tip,as shown in Fig.1(b).

Fig.1 FPPR model on various facial images
2.2 Pose pre-recognition based on geometry
The geometry algorithm is based on two attributes:symmetryand rotational invariance. Once the five facial feature points are known,the facial pose classifier is obtained.
The five facial feature points are:Ple(left eye center),Pre(right eye center),Pnt(nose tip),Plm(left mouth tip),Prm(right mouth tip).There are three triangles(T1is the red triangle;T2is the yellow triangle;T3is the blue triangle)based on the five feature points as shown in Fig.1(c).The three points:Pvd(center black point),Pvl(left black point)and Pvr(right black point)are the vertical points of each triangle,as shown in Fig.1(d).
Pvdis on vectorand satisfies(1).

S1is the distance between the facial feature point Plmand the vertical point Pvd,and S2is the distance between Prmand Pvd.One of the parameters of the pre-recognition model,kdpcan be calculated by

which shows the facial symmetry situation of the triangle
Pvlis on vector,and satisfies

Pvris on vectorand satisfies(5).

S3is the distance between the nose tip Pntand the vertical point Pvl,also known as the length between the Pntand Pvr.kcpis another parameter of the pre-recognition model,and kcpis always greater than or equal to 1.The parameter kcp(kcp?1)reflects the facial symmetry situation of the triangles T2and T3.

where n1and n2are the weights of the classifier,and K is the whole parameter of the FPPR model based on the symmetry attribute.
Parameters n1and n2are proposed to help the prerecognition model to obtain a better classification rate.We take the initial weight value(n1=n2=0.5)in this paper.We have fine-tuned the value of these parameters(n1,n2).However,the recognition effect is overall improved.The distances Si(i=1,2)are almost equal and this situation is similar to Si(i=3,4)by the symmetry attribute of three triangles while the facial pose of the test image contains a complete face. The parameters kdpand kcpare close to 1.If the test face image is a profile,the distance is not closely equal.Therefore,the parameters kdpand kcpare not close to 1.The pre-recognition model also possesses the rotational invariance attribute since the full face poses kdpand kcpare close to 1 even if the test face image has been rotated.Table 1 shows the detail data of the FPPR model,and the test samples are shown in Fig.1.We can see the parameters of the full-face image are close to 1,and the parameters of the profile are not.Therefore,the threshold kthas been set for classifying the various facial poses.

Table 1 Detail data of the FPPR model from Fig.1
3.Dual-dictionary SRC
The traditional SRC method faces many challenges when the test samples are affected by various poses.The reason is that different facial poses affect each other in the dictionary,even in the same class.Therefore,the algorithm is named dual-dictionary SRC(DD-SRC),since it contains two dictionaries,a full face dictionary and a profile one.For atoms of each dictionary,the more similar they are in a dictionary,the better,since it would reduce the influence from each other.This is one obvious defect of traditional single dictionary methods,and our algorithm is able to overcome it,in which dual-dictionary plays a fatal role.Firstly,the test face image is classified by the pre-recognition model to identify the facial poses,the full frontview or the profile view.One dictionary,full face dictionary or the profile dictionary is chosen to solve the un-derdetermined equation Ax=b based on the view of the test image.
3.1 Traditional SRC
Now,let us briefly review the procedure of the SRC algorithm.Suppose there are M classes and N training samples.The one-dimensional column vectors of all the N training samples are denoted as v1...vN.y represents the one-dimensional form of the test sample.All of the N training samples are denoted as the training matrix D=[v1,v2,...,vN].The SRC algorithm assumes that the test sample y can be adequately represented by the sparse linear combination of training samples.Especially,the purpose of the SRC algorithm is to obtain x which satisfies the following objective function:

where??1represents l1-norm of a vector,and x corresponds to the coefficient column vector.The algorithm utilizes an iterative method to gain the optimal x =[a1,a2,...,aN],which converts the problem mentioned above into the following optimization problem:

where??2denotes the l2-norm of a vector,and?y-Dixi?22is the residual of the ith class.Diis the training dictionary matrix that is composed of all training samples from the ith class and xiis a vector that consists of the weights of training samples.The SRC algorithm finds the sparsest solution and that class of the smallest residual coefficient can be consideredas the same class of y[12–14].
3.2 Full face dictionary
Enough information of facial features for each class is included in one full face dictionary.Therefore,we use the face images directly without any processing to build up the full face dictionary

where Dfis the full face dictionary,which contains k classes,and each class containsn atoms.Airepresents the ith class of Dfand vi,jrepresents the jth atom in Ai.The given test facial sample,well aligned,can be known as one of the linear combination of the k·n-dimensional column vectors,while it is recognized as a full face by the pre-recognition model

where aiis the corresponding coefficient of vi,j,and the aim is to find the sparsest solution.Then the class of the largest coefficient can be considered as the same class as yf.In this way,we can obtain the classification result of a full face.
3.3 Horizontal symmetry extension
The purpose of the horizontal symmetry extension is to obtain the detail facial information of the profile.The horizontal symmetry extension is utilized on the atoms of the dictionary while the training facial image is profile.There are two steps of the horizontal symmetry extension.
In the first step,

Suppose the size of the facial image f(x,y)is m×n,the size of result fh(x,y)will be m×2n,as shown in Fig.2(a)and Fig.2(b).The first row pictures are included in Fig.2(a).The second and third row pictures are included in Fig.2(b).
In the second step,

A test facial image is used to demonstrate this process.Details of horizontal symmetry extension are shown in Fig.2.

Fig.2 Horizontal symmetry extension on various facial images
3.4 profile dictionary
The profile dictionary contains various poses of facial images,and the influence of the full face is reduced largely which leads to better recognition performance.
Two different poses are obtained by the symmetry extension algorithm that means the information from the limited training samples can be mined for building the profile training dictionary.

where Dpis the profile dictionary,Aiis the i th class of the Dpandis the result of horizontal symmetry extension of Ai.
Processed by the symmetry extension,the given profile test facial image can be seen as one of the linear combination of the 2kn-dimensional column vectors

where aiis the corresponding coefficient of vi,j,and the class of the largest coefficient can be considered as the same class as yp.The vi,jandare in the same class.In the end,the classification result is based on the smallest residual of the solution of(10).
4.Experimental results
In this section,our approach is utilized on ORL database and FERET database for face recognition.We test the face recognition rates of DD-SRC,compared with the traditional single dictionary SRC algorithm.DD-SRC includes two facial posture dictionaries,a full face dictionary and a profile dictionary.The experimenter will divide the facial posture of the training sample into two categories:full face and profile.If the number of training samples in a certain posture dictionary(such as a full face dictionary)is insufficient,the training sample of the full face dictionary is selected from the opposing posture(profile dictionary)to ensure that the number of training samples for each type of postures is consistent.
4.1 Performance on ORL database
The ORL database includes400 frontal images for 40 individuals.Samples of this database consist of facial variation like various postures and expressions,which can be obstacles or challenges for the system to find the true class of the test image.
We test the face recognition rate of the algorithm under the condition of low training samples.In other words,for each dictionary,the facial information is not sufficient especially for various facial poses.For the construction dictionary A,we try to reduce the number of training samples(W =1,2 or 3 samples per class),which means the dimension of each training dictionary is 40,80 or 120,and the remaining samples can be used as the test samples.All these images have been simply down sampled.Threshold ktis the empirical value,and the value is set to 1.50.The five facial feature points detection is just for the pre-recognition,and it is not used for choosing atoms of dictionary.To be fair,the number of base samples to construct each training dictionary is the same in comparison of various classification algorithms.
For DD-SRC,we choose the full face poses and proifle images for the corresponding dictionary(if some class has no corresponding image,the inverse pose image will be chosen),and we run it five times to calculate the average recognition rate.All results of the proposed algorithm under each condition are shown in Table 2.Besides,our proposed algorithm is tested,and several stateof-art face recognition algorithms such as SRC[15],two phase test sample sparse representation(TPTSSR)[16],the feature space-based representation algorithm proposed in[17],symmetrical face based two-step face recognition[10],and the linear regression classification(LRC)[18]are compared.All the results are shown in Table 3.
In Table 3,although the face recognition rate under the condition of W=2 is not the highest of all compared algorithms,the face recognitionrate under W=i(i=1,3)conditions is higher than other algorithms.

Table 2 The face recognition rate of our proposed algorithm under low training samples on the ORL and FERET databases %

Table 3 Results of comparison on the ORL database %
4.2 Performance on FERET database
The FERET database contains different lighting variations,facial expressions and poses.Compared with other face databases,the FERET database has more various facial poses.Each individual has seven different facial images.Similar experiments are performed on the FERET database.All facial images have also been simply down sampled.The value of threshold ktis set to an experience value 1.59.In each dictionary of DD-SRC,we set three cases(W=1,2 and 3),which means the number of the training sample per class of each pose dictionary is 1,2and 3.For the DD-SRC,there are two poses dictionaries,profile dictionary and full face dictionary,and the experiments are repeated five times to get the average face recognition rate.
We choose the full face pose and profile images for the corresponding dictionary.All results of the experiments described above are shown in Table 2.
Table 4 shows the results of the comparison of several algorithms on the FERET database.In Table 4,although the face recognition rate when W=1 is almost equal to the symmetrical face based two-step face recognition algorithm,and the recognition rate under W=2 just has a tiny improvement.When W=3,our algorithm outperforms other algorithms by a large margin.Table 3 also shows a phenomenon that the face recognition is decreased from the condition of W=2 to W=3.The decrease rate of our proposed algorithm is the slowest compared with other algorithms under the condition from W=2 to W=3,since the mutual interference between different poses is reduced by FPPR and the dual-dictionary model.

Table 4 Results of comparison on the FERET database %
5.Conclusions
In this paper, we propose a significant algorithm under various poses and non-sufficient training samples.We first discuss the disadvantages of the traditional recognition algorithm SRC,which is based on the single dictionary.Then an improved method is proposed to make the recognition rate higher.This method consists of two parts,FPPR and DD-SRC.The different poses are affected by each other when they are in one dictionary.The dual-dictionary divides the various poses into two dictionaries,full face dictionary and profile dictionary.Finally,we prove this method by experiments on ORL database and FERET database.
Unfortunately,since the number of facial images in the data base is limited,the FPPR model is affected,which may lead to misclassification.When W=1 or W=2,due to the small number of training samples,the characterization of the posture is weak in DD-SRC.Therefore,the result is not as good as the compared method in those cases(ORL,W=2;FERET,W=1).
Our future work is to enhance the ability of pre-recognition to adapt to more various facial poses.Besides,we still need further study on how to build the effective training dictionary.
[1]ZHANG L,CHEN S,QIAO L.Graph optimization for dimensionality reduction with sparsity constraints.Pattern Recognition,2012,45(3):1205–1210.
[2]LIU B D,SHEN B,GUI L,et al.Face recognition using class specific dictionary learning for sparse representation and collaborative representation.Neurocomputing,2016,204(6):198–210.
[3]PRABHU U,HEO J,SAVVIDES M.Unconstrained pose-invariant face recognition using 3D generic elastic models.IEEE Trans.on Pattern Analysis&Machine Intelligence,2011,33(10):1952–1961.
[4]LI H,SUEN C Y.Robust face recognition based on dynamic rank representation.Pattern Recognition,2016,60(12):13–24.
[5]ZHANG X Z,GAO Y S.Face recognition across pose:a review.Pattern Recognition,2009,42(11):2876–2896.
[6]TAN X Y,CHEN SC,ZHOU ZH,et al.Face recognition from a single image per person:a survey.Pattern Recognition,2006,39(9):1725–1745.
[7]TAN X,CHEN S,ZHOU Z H,et al.Recognizing partially occluded,expression variant faces from single training image per person with SOM and soft k-NN ensemble.IEEE Trans.on Neural Networks,2005,16(4):875–886.
[8]RYU Y S,OH S Y.Simple hybrid classifier for face recognition with adaptively generated virtual data.Pattern Recognition Letters,2002,23(7):833–841.
[9]QIAO L,CHEN S,TAN X.Sparsity preserving discriminant analysis for single training image face recognition.Pattern Recognition Letters,2010,31(5):422–429.
[10]XU Y,ZHU X J.Using the original and ‘symmetrical face’training samples to perform representation based two-step face recognition.Pattern Recognition,2013,46(4):1151–1158.
[11]SUN Y,WANG X G,TANG X O.Deep convolutional network cascade for facial point detection.Proc.of theIEEE Conference on Computer Vision and Pattern Recognition,2013:3476–3483.
[12]WRIGHT J,MA Y,MAIRAL J,et al.Sparse representation for computer vision and pattern recognition.Proc.of the IEEE,2010,98(5):1031–1044.
[13]ZHAO J,HUANG L X,JIA J,et al.A fast iterative pursuit algorithm in robust face recognition based on sparse representation.Mathematical Problems in Engineering,2014,2014(2):1–11.
[14]ZHANG G Q,SUN H J,JI Z X,et al.Cost-sensitive dictionary learning for face recognition.Pattern Recognition,2016,60:613–629.
[15]KOH K,KIM S,BOYD S.A matlab solver for largescale?1-regularized least squares problems. http://www.stanford.edu/~boyd/l1 ls/.
[16]XU Y,ZHANG D,YANG J,et al.A two-phase test sample sparse representation method for use with face recognition.IEEE Trans.on Circuits and Systems for Video Technology,2011,21(9):1255–1262.
[17]XU Y,FAN Z,ZHU Q.Feature space-based human face image representation and recognition.Optical Engineering,2012,51(1):017205.
[18]NASEEM I,TOGNERI R,BENNAMOUN M.Linear regression for face recognition.IEEE Trans.on Pattern Analysis and Machine Intelligence,2010,32(11):2106–2112.
杂志排行
Journal of Systems Engineering and Electronics的其它文章
- t/k-fault diagnosis algorithm of n-dimensional hypercube network based on the MM∗model
- Joint optimization of sampling interval and control for condition-based maintenance using availability maximization criterion
- Improved evidential fuzzy c-means method
- Event-triggered sampling and fault-tolerant control co-design based on fault diagnosis observer
- Improved quantum bacterial foraging algorithm for tuning parameters of fractional-order PID controller
- Controller design for stochastic nonlinear systems with matched conditions
