CN106778522B - Single-sample face recognition method based on Gabor feature extraction and spatial transformation - Google Patents

Single-sample face recognition method based on Gabor feature extraction and spatial transformation Download PDF

Info

Publication number
CN106778522B
CN106778522B CN201611059543.6A CN201611059543A CN106778522B CN 106778522 B CN106778522 B CN 106778522B CN 201611059543 A CN201611059543 A CN 201611059543A CN 106778522 B CN106778522 B CN 106778522B
Authority
CN
China
Prior art keywords
gabor
matrix
feature
transformation
class
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Active
Application number
CN201611059543.6A
Other languages
Chinese (zh)
Other versions
CN106778522A (en
Inventor
葛洪伟
李莉
江明
朱嘉钢
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Shenzhen Huiyouba Technology Co.,Ltd.
Original Assignee
Jiangnan University
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Jiangnan University filed Critical Jiangnan University
Priority to CN201611059543.6A priority Critical patent/CN106778522B/en
Publication of CN106778522A publication Critical patent/CN106778522A/en
Application granted granted Critical
Publication of CN106778522B publication Critical patent/CN106778522B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
    • G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
    • G06V40/16—Human faces, e.g. facial parts, sketches or expressions
    • G06V40/168—Feature extraction; Face representation
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
    • G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
    • G06V40/16—Human faces, e.g. facial parts, sketches or expressions
    • G06V40/172—Classification, e.g. identification

Landscapes

  • Engineering & Computer Science (AREA)
  • Health & Medical Sciences (AREA)
  • Oral & Maxillofacial Surgery (AREA)
  • General Health & Medical Sciences (AREA)
  • Human Computer Interaction (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Multimedia (AREA)
  • Theoretical Computer Science (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Image Analysis (AREA)

Abstract

The invention discloses a single-sample face recognition method based on Gabor feature extraction and spatial transformation, which mainly solves the problem that the traditional face recognition method cannot be applied due to the fact that an intra-class scattering matrix is zero under the condition that only a single training sample image exists. The method comprises the steps of extracting a spatial feature vector from an original single-sample image by adopting Gabor wavelet, fusing the extracted spatial feature vector and an original spectral feature vector, performing low-dimensional feature space transformation on a fusion feature matrix by using a feature space transformation method, transforming the fusion feature matrix into a low-dimensional subspace, and finally completing identification by using a nearest neighbor classifier. The method can accurately finish the recognition of the single sample face, improve the recognition precision and reduce the calculation cost. Compared with the prior art, the face recognition method provided by the invention has higher effectiveness and robustness.

Description

Single-sample face recognition method based on Gabor feature extraction and spatial transformation
Technical Field
The invention belongs to the field of pattern recognition and image processing, and relates to a face recognition problem under the condition that a traditional face recognition method cannot be used for a single face image due to the fact that an intra-class scattering matrix is zero; in particular to a single sample face recognition method based on Gabor feature extraction and space transformation, which can be used for video monitoring, identity recognition and the like under the single sample situation.
Background
The face recognition technology is the most important one in the biological feature recognition technology, and is widely applied in the fields of video monitoring, supervision and law enforcement, multimedia, process control, identity recognition and the like at present. Many researchers have made much research in this regard to date. However, in severe or specific environments, a completely new challenge is often posed to a face recognition system, for example, law enforcement officers only have one face image on a criminal identity card, and only can monitor and compare the face image. For such a scenario with only one face image, the face recognition problem will become very difficult, mainly because the intra-class dispersion matrix in the commonly used classification model is zero, and the traditional methods such as Fisher linear discriminant analysis, maximum divergence difference, etc. cannot be directly used. This situation we often call the single training sample face recognition problem in an unconstrained environment. How to complete accurate automatic identification by monitoring and capturing a single brand new face of a criminal under the conditions of poor illumination, large change of facial posture and large change of expression is a great challenge. At present, the problem of face recognition of a single training sample is not well solved.
In recent years, some researchers at home and abroad have conducted some research on single-sample training images, Gao et al proposed Fisher Linear discriminant analysis (Gao Q, Zhang 567 registration using F L DA with single recording image person [ J ]. Applied Mathesics and computing, 2008, 205:. 726-.
Although both SVD-based and QRCP-based Fisher linear discriminant analysis methods can solve the problem of face recognition of a single training sample, the two methods have the following three disadvantages: (1) the reconstructed approximate image is not very satisfactory or convincing; (2) in the QRCP-based method, no theoretical analysis and interpretation is given to the approximation image to contain at least 97% of the energy of the original image; in the SVD-based method, when the number of approximate images is greater than 4, there is only a slight difference between the base image and the original image. (3) In the SVD and QRCP based methods, decomposition and storage of large scale images is omitted.
The SDD based method is superior to SVD and QRCP based methods in terms of recognition rate and recognition time, and requires less storage space, but there are three major drawbacks to the SDD based method: (1) the stopping criterion of the method needs manual control; (2) SDD-based methods still use Fisher criteria, namely: the method still utilizes the intra-class dispersion matrix and the inter-class dispersion matrix to obtain effective discrimination information.
Disclosure of Invention
Aiming at the existing problems, the invention provides a single-sample face recognition method based on Gabor feature extraction and spatial transformation, which aims to solve the problem that an intra-class scattering matrix under the single-sample image scene is zero. Therefore, the accuracy and robustness of the face recognition in practical application are improved.
The key technology for realizing the invention is as follows: under the condition of a single training sample image, firstly, extracting a spatial feature vector from an original single training sample image by utilizing a Gabor wavelet; then, obtaining a fusion feature matrix by using the extracted spatial feature vector and the original spectral feature vector; and then, performing low-dimensional feature space transformation on the fusion feature matrix by using a feature space transformation method, and transforming the fusion feature matrix into a low-dimensional subspace. And finally, identifying by using a nearest neighbor classifier. The single-sample face recognition method based on Gabor feature extraction and spatial transformation not only greatly improves the recognition rate and reduces the complexity of calculation, but also has higher effectiveness and robustness compared with the prior art.
In order to achieve the above object, the specific implementation steps are as follows:
(1) spatial information of a single image is extracted using a Gabor wavelet.
(1.1) constructing a Gabor filter function: the invention adopts a two-dimensional Gabor filter to extract the spatial information of a single image, which is a Gaussian kernel function adjusted by a complex sinusoidal plane. Is defined as:
Figure GDA0002523346630000031
where f is the central angular frequency of the complex sinusoidal plane wave, θ represents the normal parallel fringe direction of the Gabor function, φ is the phase, σ is the standard deviation, and γ is the spatial ratio for specifying the supporting ellipticity of the Gabor function.
(1.2) constructing a Gabor filter bank: since the Gabor filter bank is composed of a group of Gabor filters with different frequencies and directions, in the present invention, we use Gabor filter banks with five different scales and eight different directions, and the following two formulas give a Gabor filter bank with five different scales and eight different directions:
Figure GDA0002523346630000032
Figure GDA0002523346630000033
(1.3) Gabor representation of face image: for a face image a (x, y), its Gabor representation can be obtained by convolution of the original image and Gabor filtering, i.e.:
Figure GDA0002523346630000034
g (x, y) represents the result of two-dimensional convolution of Gabor filter in different scales u and different directions v, and the size of G (x, y) is determined by the down-sampling factor ξ, and the zero mean and unit variance are performed to the G (x, y) to obtain a filter characteristic matrix Zu,v∈Rm*n。
(1.4) constructing a Gabor direction block feature matrix: filtering characteristic matrix Z obtained in (1.3)u,vConversion into one-dimensional column vector, using Z0Indicating surfaceThe Gabor direction block feature matrices of five different scales and eight different directions of the partial image a (x, y) are as follows:
Figure GDA0002523346630000035
wherein the content of the first and second substances,
Figure GDA0002523346630000036
is Zu,vOne-dimensional representation at the scale i, Z0∈R(m*n)×40Is a Gabor direction block feature matrix obtained from the convolution result G (x, y).
(2) And fusing the spatial information extracted by the Gabor wavelet and the spectral information of the original image.
The Gabor direction block feature matrix obtained in the step 1 is Z0∈R(m*n)×40On the other hand, since a single training sample image itself contains very important spectral information, the Gabor spatial feature vector and the spectral feature vector Y are combined0∈R(m *n)×1Performing fusion to obtain a fusion feature matrix F ∈ R(m*n)×41：
Figure GDA0002523346630000037
Wherein σ1And σ2Are each Z0And Y0Can be obtained by calculating the square root of the variance of the feature vector.
3. Feature space transformation based on the fused feature matrix.
(3.1) establishing a fusion characteristic optimization model
Figure GDA0002523346630000041
(3.2) performing feature space transformation to obtain a transformation matrix
Figure GDA0002523346630000042
Wherein the content of the first and second substances,
Figure GDA0002523346630000043
representing the inter-Gabor-class scatter matrix,
Figure GDA0002523346630000044
representing a dispersion matrix within the Gabor class,
Figure GDA0002523346630000045
is a matrix of orthogonal columns,
Figure GDA0002523346630000046
is a diagonal matrix with non-increasing and positive diagonal elements,
Figure GDA0002523346630000047
is an orthogonal matrix in which the matrix is orthogonal,
Figure GDA0002523346630000048
Figure GDA0002523346630000049
is a diagonal matrix;
4. constructing projection feature vectors
Figure GDA00025233466300000410
For test feature vector f ∈ Rn×1The projected feature vector can be obtained by linear transformation as follows
Figure GDA00025233466300000411
Figure GDA00025233466300000412
It is clear that the computational complexity described above is significantly reduced.
5. After the projected feature vectors are obtained, the nearest neighbor classifier is used for identification.
The method of the invention has the following advantages:
(1) the difficult problems encountered by a single training sample are overcome: since the intra-class dispersion matrix in the single training sample model is zero, the conventional Fisher criterion fails, and the method reconstructs the intra-class dispersion matrix through Gabor filtering and feature space transformation.
(2) The invention can make full use of the spatial information of the original image and the spectral information of the original image. Meanwhile, the spatial characteristic information based on the Gabor is more robust than the spectral characteristic information of the image, and can avoid local distortion caused by changes of expressions, postures, illumination and the like. The recognition rate and the recognition time are greatly improved, and the calculation cost is reduced. Compared with the prior art, the face recognition method provided by the invention has higher effectiveness and robustness.
Drawings
FIG. 1 is a flow chart of the method of the present invention.
Fig. 2Gabor filters the real part at 5 different scales and 8 different directions.
Fig. 3 shows the convolution result of 2Gabor filters for a single face image.
Fig. 45 different facial images from each data set.
The recognition rates of the four methods in fig. 5 (OR L face database) under different projections.
Fig. 6 shows the recognition rate of the four methods under different projections (Yale face database).
Fig. 7 shows the recognition rates (FERET face database) of the four methods in different projections.
Detailed Description
The invention relates to a single-sample face recognition method based on Gabor feature extraction and spatial transformation. Referring to fig. 1, the embodied steps of the present invention include the following.
And step 1, extracting the spatial information of a single image by using Gabor wavelet.
(1.1) constructing a Gabor filter function: the invention adopts a two-dimensional Gabor filter to extract the spatial information of a single image, which is a Gaussian kernel function adjusted by a complex sinusoidal plane. Is defined as:
Figure GDA0002523346630000051
where f is the central angular frequency of the complex sinusoidal plane wave, θ represents the normal parallel fringe direction of the Gabor function, φ is the phase, σ is the standard deviation, and γ is the spatial ratio for specifying the supporting ellipticity of the Gabor function.
(1.2) constructing a Gabor filter bank: since the Gabor filter bank is composed of a group of Gabor filters with different frequencies and directions, in the present invention, we use Gabor filter banks with five different scales and eight different directions, and the following two formulas give a Gabor filter bank with five different scales and eight different directions:
Figure GDA0002523346630000052
Figure GDA0002523346630000053
(1.3) Gabor representation of face image: for a face image a (x, y), its Gabor representation can be obtained by convolution of the original image and Gabor filtering, i.e.:
Figure GDA0002523346630000054
wherein G (x, y) represents the result of the two-dimensional convolution of the Gabor filter at different scales u and different directions v. And G (x, y)
Is determined by the downsampling factor ξ, and is zeroed to obtain a filter feature matrix Zu,v∈Rm*n。
(1.4) constructing a Gabor direction block feature matrix: filtering characteristic matrix Z obtained in (1.3)u,vConversion into one-dimensional column vector, using Z0The Gabor direction block feature matrix representing five different scales and eight different directions of the face image a (x, y) is as follows:
Figure GDA0002523346630000055
wherein the content of the first and second substances,
Figure GDA0002523346630000056
is Zu,vOne-dimensional representation at the scale i, Z0∈R(m*n)×40Is a Gabor direction block feature matrix obtained from the convolution result G (x, y).
And 2, fusing the spatial characteristic information extracted by the Gabor characteristic and the spectral information of the original image.
Since a single sample image contains very important spectral information, the Gabor spatial feature vector and the spectral feature vector Y are combined0∈R(m*n)×1Performing fusion to obtain a fusion characteristic matrix F ∈ R(m*n)×41：
Figure GDA0002523346630000057
Wherein σ1And σ2Are each Z0And Y0Can be obtained by calculating the square root of the variance of the feature vector.
And 3, performing feature space transformation on the fusion feature matrix F.
(3.1) establishing a fusion characteristic optimization model
Figure GDA0002523346630000061
In order to be able to distinguish each class of all training images, it is desirable that the differences from the same class are as small as possible, whereas the differences between sample images from different classes are as large as possible. Inspired by Fisher criterion thought, the following fusion characteristic optimization model is established:
Figure GDA0002523346630000062
our goal is to project the fused feature matrix into a low-dimensional feature subspace using feature space transformations and find an optimal linear transformation matrix that maximizes inter-class separation.
(3.2) performing feature space transformationTransforming to obtain transformation matrix W2。
(3.2a) construction of inter-Gabor-type spreading matrices and intra-Gabor-type spreading matrices.
Suppose that n-dimensional training samples are obtained from the fused feature matrix, c is the number of classes, ni(i ═ 1,2, …, c) is the number of training samples of class i, and then the inter-Gabor scattering matrix and the intra-Gabor scattering matrix are defined as follows:
Figure GDA0002523346630000063
Figure GDA0002523346630000064
wherein the content of the first and second substances,
Figure GDA0002523346630000065
is the jth fused feature vector from class i, f is the mean vector of class i, f0Is the mean vector of all training samples.
We maximize at the same time
Figure GDA0002523346630000066
And minimization of
Figure GDA0002523346630000067
A transformation matrix W is obtained. The optimization model is as follows:
Figure GDA0002523346630000068
Figure GDA0002523346630000069
but when the matrix in the above formula is
Figure GDA00025233466300000610
Sum matrix
Figure GDA00025233466300000611
The above criteria are not true when there is singularity. In this case, matrix decomposition and eigenspace transformation generally play an important role. In feature space transformation, since our goal is to maximize the diversity of different classes,
therefore, it should be discarded
Figure GDA00025233466300000612
Because it contains useless information; at the same time should remain
Figure GDA00025233466300000613
Is important information of the null space.
(3.2b) projecting Gabor inter-class spreading matrix and Gabor intra-class spreading matrix to s1Dimensional subspace and obtain a transformation W1。
In this step, first consider
Figure GDA00025233466300000614
Singular Value Decomposition (SVD)
Figure GDA0002523346630000071
Will UbThe block-shaped materials are divided into blocks,
Figure GDA0002523346630000072
wherein
Figure GDA0002523346630000073
Therefore, the temperature of the molten metal is controlled,
Figure GDA0002523346630000074
thus, an equation of the form:
Figure GDA0002523346630000075
wherein the content of the first and second substances,
Figure GDA0002523346630000076
is a matrix of orthogonal columns,
Figure GDA0002523346630000077
is a diagonal matrix with non-increasing and positive diagonal elements. In practical application, the matrix
Figure GDA0002523346630000078
The singularity of (a) may result in a reduction of the discrimination ability. Therefore, its zero eigenvalue and corresponding eigenvector should be discarded. In view of the above-mentioned considerations,
first, using transformations
Figure GDA0002523346630000079
Transforming the original data to s1In the space of the dimension. Thus obtaining the transformation
Figure GDA00025233466300000710
(3.2c) at s1Performing a correlated transformation in a transformation space of dimensions to obtain a transformation W2。
In step (3.2b) a transformation has been obtained
Figure GDA00025233466300000711
In the resulting transform space, inter-class scatter matrices
Figure GDA00025233466300000712
And intra-class scatter matrices
Figure GDA00025233466300000713
Respectively become:
Figure GDA00025233466300000714
Figure GDA00025233466300000715
wherein the content of the first and second substances,
Figure GDA00025233466300000716
the original n × n-dimensional inter-class and intra-class scatter matrices are then reduced to s1×s1And (5) maintaining.
Now, we consider
Figure GDA00025233466300000717
Decomposition of characteristic values of (2):
Figure GDA00025233466300000718
wherein
Figure GDA00025233466300000719
Is an orthogonal matrix in which the matrix is orthogonal,
Figure GDA00025233466300000720
is a diagonal matrix. Thus, there are:
Figure GDA00025233466300000721
in most of the fields of application,
Figure GDA00025233466300000722
is greater than
Figure GDA00025233466300000723
And ∑, andwis non-exotic, due to:
Figure GDA00025233466300000724
therefore, there are:
Figure GDA0002523346630000081
thus, the optimal transformation matrix
Figure GDA0002523346630000082
This can be obtained by the following formula:
Figure GDA0002523346630000083
in fact, the above optimization problem
Figure GDA0002523346630000084
The following eigenvalue problem can be transformed to solve:
Figure GDA0002523346630000085
and the solution of the above characteristic problem can be obtained by solving the generalized characteristic value. Let λ be1，λ2，…，λtT maximum eigenvalues in descending order, w, of the eigenvalue problem1，w2，…，wtIs the corresponding feature vector.
To solve the problem
Figure GDA0002523346630000086
Two steps are mainly considered: the first step is to maximize the inter-class scatter matrix by Singular Value Decomposition (SVD) and the second step is to solve the generalized eigenvalue problem. The key problem of the first step is to deal with the following optimization problem:
Figure GDA0002523346630000087
we know that:
Figure GDA0002523346630000088
is a matrix of orthogonal columns,
Figure GDA0002523346630000089
is a diagonal matrix with non-increasing and positive diagonal elements. Thus, U can be obtainedb1Is a solution to the above problem.
In addition, the pseudo-inverse is usually used to solveAnd (5) solving the singular matrix. The pseudo-inverse of the matrix may be computed by Singular Value Decomposition (SVD). One natural extension using the pseudo-inverse is to use a feature decomposition matrix
Figure GDA00025233466300000810
Or
Figure GDA00025233466300000811
More specifically, let M be μ ∑ VTIs a singular value decomposition of M, where U and V are column orthogonal matrices, ∑ is a diagonal matrix with diagonal elements positive, and then the pseudo-inverse of M is M+＝V∑-1UT. Based on the above discussion, we have obtained
Figure GDA00025233466300000812
Is optimized to transform the matrix
Figure GDA00025233466300000813
Thus, based on the above transformations and argumentations, we derive s after transformation1In the dimension space of the optical fiber,
Figure GDA00025233466300000814
is an optimal transformation matrix.
Step 4, constructing projection characteristic vector
Figure GDA00025233466300000815
For face test image I ∈ RM×NThe Gabor-based spatial feature matrix can be formulated from
Figure GDA00025233466300000816
Figure GDA00025233466300000817
Is obtained and expressed by
Figure GDA00025233466300000818
A fused feature matrix may be obtained. Therefore, a newThe Gabor direction block feature matrix of the face test image
Figure GDA00025233466300000819
By the formula
Figure GDA00025233466300000820
Figure GDA00025233466300000821
And
Figure GDA00025233466300000822
it is obtained.
And 5, finally, utilizing a nearest neighbor classifier to identify.
Based on the above description, for a face image a (x, y), spatial feature information is extracted through Gabor wavelet, then the spatial feature information is fused with the spectral information of the original image, and then the feature space transformation is performed, and the transformed s1In the dimensional subspace, we obtain the optimal transformation matrix
Figure GDA0002523346630000091
And obtains the projection feature vector
Figure GDA0002523346630000092
And then identified by a nearest neighbor classifier. The nearest neighbor classifier (NNc) is a non-parametric method classifier, and the main idea is that: let X { (X) be the training sample set1,l1),(x2,l2),…,(xn,ln) In which liI-1, 2, …, n is a class label if the test sample x and the k samples x of the training sample1,…,xkThere is a minimum distance between, then the test sample x belongs to liAnd (i ═ 1,2, …, k).
Suppose FtestIs a test image, according to the formula
Figure GDA0002523346630000093
The Gabor direction block feature matrix x can be obtainedtest. F is shown by the following formulatestBelong to the i-th class:
Figure GDA0002523346630000094
wherein the content of the first and second substances,
Figure GDA0002523346630000095
i＝1,2…,c，j＝1,2…,ni。
the effects of the present invention can be further illustrated by the following simulation experiments.
1. Simulation conditions and parameters
An example of the invention is for the OR L, Yale and FERET face databases, the OR L database contains 400 112 × 92 sized face images of 40 individuals, 10 for each, with changes in pose, angle, scale, expression, glasses, etc. since these images were taken at different times FIG. 4(a) shows 5 different face images of the present dataset, 1 image of each individual in the present experiment is used for training, the remaining 9 images are used for testing, the Yale face database contains 165 images from 15 individuals, 11 images for each, these images will change with changes in facial expression and lighting conditions, e.g., happy, sad, frightened, cold, glasses, etc. FIG. 4(b) shows 5 different face images of the present dataset, in the experiment, the size of each image is set to 100 × 100, 1 image of each individual is used for training, the remaining images are used for testing, the FET face database is created by the U.S. department through DAR's facial image project, with different face images from DAR's pose, 1195 different face images from the national defense, the present dataset, the size of each image is given by the U.S. DAR 3 facial image, the present database, the present image contains 5 different face images, the present image, the same facial image, the same, the model, the face database contains 5 face image, the.
In the experiment, we selected 5 different images of 15 persons, a total of 75 facial images, and resized the images to 80 × 80. we performed 50 experiments on the above three face databases and compared them with the existing methods, namely, the Fisher linear discriminant analysis method based on SVD, the Fisher linear discriminant analysis method based on QRCP, and the Fisher linear discriminant analysis method based on SDD for a single facial image. fig. 2 shows the Gabor filter bank for 5 different scales and 8 different directions. fig. 3 shows the convolution results of 2Gabor filters on a single facial image.
2. Simulation content and result analysis
In a simulation test, the method of the invention is compared and analyzed with the traditional Fisher linear discriminant analysis method based on SVD, the Fisher linear discriminant analysis method based on QRCP and the Fisher linear discriminant analysis method based on SDD, and the test is developed on three data sets.
Experiment one:
the experiment is implemented on an OR L face database according to the five steps, the result of the experiment about the recognition rate is shown in FIG. 5, and it can be seen from FIG. 5 that, in the four methods, the maximum recognition rate of the proposed method is 76.67%, which is higher than that of the other methods (the maximum recognition rate of the Fisher linear discriminant analysis method based on SVD is 56.67%, the maximum recognition rate of the Fisher linear discriminant analysis method based on QRCP is 68.89%, and the maximum recognition rate of the Fisher linear discriminant analysis method based on SDD is 71.94%). compared with the other three methods, the proposed method has the maximum recognition rate and obtains the best recognition performance.
Experiment two:
second experiment is implemented on the Yale face database according to the five steps, fig. 6 shows the change relationship of the recognition rate of the four methods with respect to the projection vector, and from fig. 6, we can draw the following conclusions: according to the method, the identification rate is gradually increased along with the increase of the number of the projection vectors, and the identification performance is gradually enhanced. The maximum recognition rates of the Fisher linear discriminant analysis method based on SVD, the Fisher linear discriminant analysis method based on QRCP, the Fisher linear discriminant analysis method based on SDD and the proposed methods are 24.00%, 38.67%, 45.33% and 64.67%, respectively.
From the recognition rate, the method is optimal, the two methods of the Fisher linear discriminant analysis method based on the SDD and the Fisher linear discriminant analysis method based on the QRCP are suboptimal, and the recognition performance of the Fisher linear discriminant analysis method based on the SVD is the worst, which can be fully illustrated in FIG. 6
Experiment three:
experiment three was performed on the FERET face database, again in the five steps described above. The Fisher linear discriminant analysis method based on SVD, the Fisher linear discriminant analysis method based on QRCP, the Fisher linear discriminant analysis method based on SDD and the relation between the recognition rate and the projection vector of the methods are shown in FIG. 7. As is clear from fig. 7, the recognition performance of the proposed method is higher than that of the other three methods, and the recognition rate thereof gradually increases as the number of projection vectors increases, which shows excellent recognition performance. The best recognition rate of the Fisher linear discriminant analysis method based on SVD, the Fisher linear discriminant analysis method based on QRCP, the Fisher linear discriminant analysis method based on SDD and the provided method is respectively as follows: 88.83%, 86.67%, 93.33% and 96.67%.
The three experiments show that the single-sample face recognition method based on Gabor feature extraction and spatial transformation has relatively good recognition results in actual face recognition because Gabor feature blocks are robust to local distortion caused by changes of expression, posture and illumination, and Table 1 shows the maximum recognition rate (rr,%) and recognition time (t, s) #1, #2 and #3 of the four methods on three different data sets respectively represent OR L, Yale and FERET data sets.
Table 1 maximum recognition rate (rr,%) and recognition time (t, s) for the four methods on different datasets.
Figure GDA0002523346630000111
As can be seen from Table 1, the recognition rate of the proposed method is higher than that of the other three methods (i.e., Fisher Linear discriminant analysis method based on SVD, Fisher Linear discriminant analysis method based on QRCP, Fisher Linear discriminant analysis method based on SDD) and is much shorter in recognition time than the other three methods.
It is obvious from the experimental result chart that the recognition rate of the method of the invention is obviously higher than that of the Fisher linear discriminant analysis method based on SVD, the Fisher linear discriminant analysis method based on QRCP and the Fisher linear discriminant analysis method based on SDD, and the average recognition time is obviously lower than that of other three algorithms. This temporal difference is mainly caused by image vectorization. Therefore, the face recognition method is a very effective single-sample face recognition method with good robustness.

Claims (3)

1. A single-sample face recognition method based on Gabor feature extraction and spatial transformation comprises the following steps:
(1) extracting spatial information of a single image by using a Gabor wavelet:
(1.1) constructing a Gabor filter function: extracting spatial information of a single image by adopting two-dimensional Gabor filtering, wherein the spatial information is a Gaussian kernel function adjusted by a complex sinusoidal plane; is defined as:
Figure FDA0002523346620000011
wherein f is the central angular frequency of the complex sinusoidal plane wave, θ represents the normal parallel stripe direction of the Gabor function, φ is the phase, σ is the standard deviation, γ is the spatial ratio for specifying the supporting ellipse of the Gabor function;
(1.2) constructing a Gabor filter bank: since the Gabor filter bank is composed of a set of Gabor filters with different frequencies and directions, using a Gabor filter bank with five different scales and eight different directions, the following two equations give a Gabor filter bank with five different scales and eight different directions:
Figure FDA0002523346620000012
Figure FDA0002523346620000013
(1.3) Gabor representation of face image: for a face image a (x, y), its Gabor representation can be obtained by convolution of the original image and Gabor filtering, i.e.:
Figure FDA0002523346620000014
wherein G (x, y) represents the two-dimensional convolution result of Gabor filtering in different scales u and different directions v, and the size of G (x, y) is determined by a down-sampling factor ξ, and the zero mean and unit variance are carried out on the G (x, y) to obtain a filtering feature matrix Zu,v∈Rm*n；
(1.4) constructing a Gabor direction block feature matrix: filtering characteristic matrix Z obtained in (1.3)u,vConversion into one-dimensional column vector, using Z0The Gabor direction block feature matrix representing five different scales and eight different directions of the face image a (x, y) is as follows:
Figure FDA0002523346620000015
wherein the content of the first and second substances,
Figure FDA0002523346620000016
is Zu,vOne-dimensional representation at the scale i, Z0∈R(m *n)×40Is a Gabor direction block feature matrix obtained from the convolution result G (x, y);
(2) and fusing spatial information extracted by Gabor wavelet and spectral information of the original image: obtaining a Gabor direction block feature matrix Z from the step (1)0∈R(m*n)×40On the other hand, since a single training sample image itself contains very important spectral information, the Gabor spatial feature vector and the spectral feature vector Y are combined0∈R(m*n)×1Performing fusion to obtain a fusion feature matrix F ∈ R(m*n)×41：
Figure FDA0002523346620000021
Wherein σ1And σ2Are each Z0And Y0The standard deviation of (a), which can be obtained by calculating the square root of the variance of the feature vector;
(3) feature space transformation based on the fusion feature matrix:
(3.1) establishing a fusion characteristic optimization model
Figure FDA0002523346620000022
(3.2) performing feature space transformation to obtain a transformation matrix
Figure FDA0002523346620000023
Wherein the content of the first and second substances,
Figure FDA0002523346620000024
representing the inter-Gabor-class scatter matrix,
Figure FDA0002523346620000025
representing a dispersion matrix within the Gabor class,
Figure FDA0002523346620000026
is a matrix of orthogonal columns,
Figure FDA0002523346620000027
is a diagonal matrix with non-increasing and positive diagonal elements,
Figure FDA0002523346620000028
is an orthogonal matrix in which the matrix is orthogonal,
Figure FDA0002523346620000029
Figure FDA00025233466200000210
is aA diagonal matrix;
(4) constructing a projection feature vector;
(5) after the projected feature vectors are obtained, the nearest neighbor classifier is used for identification.
2. The single-sample face recognition method based on Gabor feature extraction and spatial transformation of claim 1, wherein: the specific process of the feature space transformation described in the step (3.2) is as follows:
(3.2a) defining Gabor inter-class spreading matrices and intra-class spreading matrices:
suppose that n-dimensional training samples are obtained from the fused feature matrix, c is the number of classes, ni(i ═ 1,2, …, c) is the number of training samples of class i, and then the inter-Gabor scattering matrix and the intra-Gabor scattering matrix are defined as follows:
Figure FDA00025233466200000211
Figure FDA00025233466200000212
wherein the content of the first and second substances,
Figure FDA00025233466200000213
is the jth fused feature vector, f, from class iiIs the mean vector of class i, f0Is the mean vector of all training samples;
(3.2b) transforming the inter-class and intra-class scatter matrices to s1In the space of dimensions and obtaining a transformation W1：
First consider
Figure FDA00025233466200000214
Singular Value Decomposition (SVD)
Figure FDA00025233466200000215
Will UbThe block-shaped materials are divided into blocks,
Figure FDA00025233466200000216
wherein
Figure FDA00025233466200000217
Therefore, the temperature of the molten metal is controlled,
Figure FDA00025233466200000218
thus, the formula
Figure FDA00025233466200000219
Can be converted into the following forms:
Figure FDA0002523346620000031
wherein the content of the first and second substances,
Figure FDA0002523346620000032
is a matrix of orthogonal columns,
Figure FDA0002523346620000033
is a diagonal matrix with non-increasing and positive diagonal elements; in practical application, the matrix
Figure FDA0002523346620000034
The singularities of (a) result in a reduction of the discrimination power, and therefore their zero eigenvalues and corresponding eigenvectors should be discarded, based on the above considerations, using the transformation
Figure FDA0002523346620000035
Transforming the original data to s1In the space of the dimension;
(3.2c) at s1Performing a correlated transformation in a transformation space of dimensions to obtain a final transformation W2：
In obtainingS of1Inter-class scatter matrices in dimension transform space
Figure FDA0002523346620000036
The following steps are changed:
Figure FDA0002523346620000037
intra-class scatter matrix
Figure FDA0002523346620000038
The following steps are changed:
Figure FDA0002523346620000039
wherein the content of the first and second substances,
Figure FDA00025233466200000310
now, consider
Figure FDA00025233466200000311
Decomposition of characteristic values of (2):
Figure FDA00025233466200000312
wherein
Figure FDA00025233466200000313
Is an orthogonal matrix in which the matrix is orthogonal,
Figure FDA00025233466200000314
is a diagonal matrix; thus, there are:
Figure FDA00025233466200000315
in most of the fields of application,
Figure FDA00025233466200000316
is greater than
Figure FDA00025233466200000317
And ∑, andwis non-exotic, due to:
Figure FDA00025233466200000318
therefore, there are:
Figure FDA00025233466200000319
thus, the optimal transformation matrix
Figure FDA00025233466200000320
Can be obtained by the following formula:
Figure FDA00025233466200000321
3. the single-sample face recognition method based on Gabor feature extraction and spatial transformation of claim 1, wherein:
the specific method for constructing the projection feature vector in the step (4) comprises the following steps:
for test feature vector f ∈ Rn×1The projected feature vector can be obtained by linear transformation as follows
Figure FDA00025233466200000322
Figure FDA0002523346620000041
It is clear that the computational complexity is significantly reduced, for the facial test image I ∈ RM×NThe Gabor-based spatial feature matrix can be formulated from
Figure FDA0002523346620000042
Is obtained and expressed by
Figure FDA0002523346620000043
Figure FDA0002523346620000044
A fusion feature matrix can be obtained;
thus, a new Gabor directional block feature matrix for the face test image
Figure FDA0002523346620000045
This can be obtained from the following equation:
Figure FDA0002523346620000046
Figure FDA0002523346620000047
Figure FDA0002523346620000048
Figure FDA0002523346620000049
then x is the required Gabor directional block feature matrix, here denoted as
Figure FDA00025233466200000410
CN201611059543.6A 2016-11-25 2016-11-25 Single-sample face recognition method based on Gabor feature extraction and spatial transformation Active CN106778522B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201611059543.6A CN106778522B (en) 2016-11-25 2016-11-25 Single-sample face recognition method based on Gabor feature extraction and spatial transformation

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201611059543.6A CN106778522B (en) 2016-11-25 2016-11-25 Single-sample face recognition method based on Gabor feature extraction and spatial transformation

Publications (2)

Publication Number Publication Date
CN106778522A CN106778522A (en) 2017-05-31
CN106778522B true CN106778522B (en) 2020-08-04

Family

ID=58911568

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201611059543.6A Active CN106778522B (en) 2016-11-25 2016-11-25 Single-sample face recognition method based on Gabor feature extraction and spatial transformation

Country Status (1)

Country Link
CN (1) CN106778522B (en)

Families Citing this family (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107403194B (en) * 2017-07-26 2020-12-18 广州慧扬健康科技有限公司 Skin cancer image recognition visualization system based on t-SNE
CN107798308B (en) * 2017-11-09 2020-09-22 一石数字技术成都有限公司 Face recognition method based on short video training method
CN107886090B (en) * 2017-12-15 2021-07-30 苏州大学 Single-sample face recognition method, system, equipment and readable storage medium
CN114445720B (en) * 2021-12-06 2023-06-20 西安电子科技大学 Hyperspectral anomaly detection method based on spatial spectrum depth synergy

Family Cites Families (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
KR101743927B1 (en) * 2010-09-09 2017-06-20 삼성전자주식회사 Method and apparatus for generating an objected descriptor using extended curvature gabor filter
CN102855468B (en) * 2012-07-31 2016-06-29 东南大学 A kind of single sample face recognition method in photograph identification
CN103514443B (en) * 2013-10-15 2016-12-07 中国矿业大学 A kind of single sample recognition of face transfer learning method based on LPP feature extraction
WO2015074157A1 (en) * 2013-11-25 2015-05-28 Ehsan Fazl Ersi System and method for face recognition
CN104239856B (en) * 2014-09-04 2017-10-17 电子科技大学 Face identification method based on Gabor characteristic and self adaptable linear regression

Also Published As

Publication number Publication date
CN106778522A (en) 2017-05-31

Similar Documents

Publication Publication Date Title
Li et al. Overview of principal component analysis algorithm
Lyons et al. Classifying facial attributes using a 2-d gabor wavelet representation and discriminant analysis
Lyons et al. Automatic classification of single facial images
CN106778522B (en) Single-sample face recognition method based on Gabor feature extraction and spatial transformation
Qian et al. Discriminative histograms of local dominant orientation (D-HLDO) for biometric image feature extraction
Timotius et al. Face recognition between two person using kernel principal component analysis and support vector machines
CN109241813B (en) Non-constrained face image dimension reduction method based on discrimination sparse preservation embedding
CN107194314B (en) Face recognition method fusing fuzzy 2DPCA and fuzzy 2DLDA
KR100345245B1 (en) Method of Processing Face Recognition Robust to Illumination Change, Facial Expression or Eyewear
CN111723759B (en) Unconstrained face recognition method based on weighted tensor sparse graph mapping
CN108664941B (en) Nuclear sparse description face recognition method based on geodesic mapping analysis
Wan et al. Face recognition with convolutional neural networks and subspace learning
Kaur et al. Comparative study of facial expression recognition techniques
CN111259780A (en) Single-sample face recognition method based on block linear reconstruction discriminant analysis
Guo et al. Palmprint Recognition Based on Local Fisher Discriminant Analysis.
De la Torre et al. Filtered component analysis to increase robustness to local minima in appearance models
Chater et al. Comparison of robust methods for extracting descriptors and facial matching
CN109919056B (en) Face recognition method based on discriminant principal component analysis
Gatto et al. Kernel two dimensional subspace for image set classification
Zhao et al. 3D object recognition and pose estimation using kernel PCA
Najim et al. Face recognition using various feature extraction approaches
Buciu et al. Non-negative matrix factorization methods for face recognition under extreme lighting variations
Vamsee Mohan et al. A Method for Face Recognition Based on Rotational Invariant Structural Texture Primitives (RISTP) Derived on Local Directional Pattern (LDP)
Bernardo et al. Exploring the scalability of multiple signatures in iris recognition using GA on the acceptance frontier search
Shen et al. A near-infrared face detection and recognition system using ASM and PCA+ LDA

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant
TR01 Transfer of patent right

Effective date of registration: 20210923

Address after: 710016 B 1616, Tiandi Times Plaza, Fengcheng two road, Weiyang District, Xi'an, Shaanxi.

Patentee after: Liu Jiaojiao

Address before: No. 1800 road 214122 Jiangsu Lihu Binhu District City of Wuxi Province

Patentee before: Jiangnan University

TR01 Transfer of patent right
TR01 Transfer of patent right

Effective date of registration: 20221026

Address after: No.06, 20th Floor, Building C, Huihai Square, Chuangye Road, Longhua Street, Longhua District, Shenzhen, Guangdong 518109

Patentee after: SHENZHEN QIANKEDUO INFORMATION TECHNOLOGY CO.,LTD.

Address before: 710016 B 1616, Tiandi Times Plaza, Fengcheng two road, Weiyang District, Xi'an, Shaanxi.

Patentee before: Liu Jiaojiao

TR01 Transfer of patent right
TR01 Transfer of patent right

Effective date of registration: 20231220

Address after: W338, 3rd Floor, Port Building, Shipping Center, No. 1167 Yihai Avenue, Nanshan Street, Qianhai Shenzhen Hong Kong Cooperation Zone, Shenzhen, Guangdong Province, 518000

Patentee after: Shenzhen Huiyouba Technology Co.,Ltd.

Address before: No.06, 20th Floor, Building C, Huihai Square, Chuangye Road, Longhua Street, Longhua District, Shenzhen, Guangdong 518109

Patentee before: SHENZHEN QIANKEDUO INFORMATION TECHNOLOGY CO.,LTD.

TR01 Transfer of patent right