Background
Face recognition is one of the most important research topics in the field of computer vision. At present, a High-Resolution (HR) face recognition method under a controlled condition tends to be mature, and starts to be popularized and applied in a large range in many production practices. However, under the actual uncontrolled condition, the performance of the face recognition system is drastically reduced under the influence of adverse factors such as posture, illumination, expression, occlusion, resolution, and the like, and cannot meet the requirements of actual application. Therefore, the recognition technology for studying Low-Resolution (LR) faces is receiving wide attention from researchers.
Over the past several decades, many different low resolution face recognition methods have been proposed. Depending on the recognition principle, there are three general categories: the method comprises an LR face recognition method based on a reconstructed Super-Resolution (SR) image, an LR face recognition method based on a public feature subspace and an LR face recognition method based on deep learning.
LR face recognition methods based on reconstructed SR images develop rapidly, and the methods mainly utilize an image SR reconstruction technology to obtain HR face images with good visual effects and achieve similarity matching of faces. Although the method based on the image SR can obtain the HR face image with higher visual effect, artifacts are easily introduced at key feature points of the face, and the recognition performance is seriously influenced; moreover, with the large-area coverage of the monitoring network, the calculation complexity of the method is high, and the actual application requirements are difficult to meet.
In recent years, the LR face recognition method based on the public feature subspace becomes an effective way for solving the problem of unmatched feature dimensions of the HR-LR face image due to the advantages of relatively simple algorithm, less time consumption and the like. The method firstly maps HR-LR face images with different dimensions to a public characteristic subspace by learning the coupling mapping of the HR-LR face, and then completes the similarity matching of the HR-LR face images in the characteristic subspace with the same dimension. At present, two common solutions are mainly provided for the problem of an LR face in a public feature subspace, wherein the first solution is an LR face recognition method based on dictionary learning and sparse representation, and the method is mainly used for performing sparse coding on local structural features of the face through the dictionary learning and the sparse representation and then transforming the face into a low-dimensional feature space to realize the matching of the LR face. The second is an LR face recognition method based on coupling mapping, which generally has 3 mapping modes: 1) sampling an HR face image to the same characteristic dimension as that of an LR face image for matching; 2) the LR face image is up-sampled to the same characteristic dimension as the HR face image for matching; 3) and meanwhile, the HR-LR face image is mapped to a common characteristic subspace for matching. The method aims to transform the HR-LR training face image features to a public feature subspace to learn an HR-LR coupling mapping matrix, and then transform the HR coupling mapping matrix and the LR coupling mapping matrix to the public feature subspace respectively to realize the transformation and identification of the LR testing face image features.
With the rapid development of deep learning, the LR face recognition method based on deep learning is proposed in succession, and compared with the traditional machine learning algorithm, the deep learning is more advantageous when a large number of training samples are processed. The face features are extracted mainly through a convolutional neural network, and effective activation functions and loss functions are adopted to optimize network parameters, so that the recognition of an end-to-end HR-LR face is realized.
Disclosure of Invention
The invention aims to provide a low-resolution face recognition method based on multi-manifold coupling mapping, which utilizes local manifold geometric structure information and label information of an HR-LR face image, enhances the discrimination capability and separability of a coupling mapping relation matrix, and improves the recognition performance of an LR face.
The technical scheme adopted by the invention is that a low-resolution face identification method based on multi-manifold coupling mapping is implemented according to the following steps:
step 1, selecting N HR facial images from a standard facial database to form an HR image set, and randomly selecting N HR facial images from the HR image set
tThe method comprises the steps of taking half of facial images of each person as an HR training set, carrying out smooth downsampling on the HR training set to generate an LR training set, and constructing class labels of training facial image samples, wherein,
step 2, based on a coupling mapping learning method, simultaneously mapping the face images in the HR training set and the LR training set to a public characteristic subspace to obtain a formula based on coupling mapping and perform matrixing;
step 3, adding the local geometric structure information and the discrimination information of the sample into the formula subjected to matrixing in the step 2, and solving an HR coupling mapping matrix PHAnd LR coupling mapping matrix PL;
Step 4, the other half of the face images in the HR image set are subjected to smooth downsampling to generate an LR test set, and the total number of the images in the test set is Np,
And 5, transforming the HR image set and the LR test set to a public feature subspace to obtain an HR-LR face mapping feature
And
step 6, applying the nearest neighbor classifier to the HR-LR face mapping feature projected in the public feature subspace
Classifying to obtain the face mapping characteristics
The category label of (1).
The present invention is also characterized in that,
the
step 1 specifically comprises the following steps: selecting N HR facial images from standard facial database to form HR image set
Randomly selecting a half of face images containing each person from an HR image set as an HR training set:
and (3) performing smooth downsampling on the HR training set to generate an LR training set:
wherein
Representing the ith low-resolution face image,
showing the ith high-resolution face image,
representing the total number of the images of the training set;
the class labels of the training face image samples are as follows:
the standard face database comprises an FERET and CMU PIE face database, and the generation of the LR training set by smoothly downsampling the HR training set specifically comprises the following steps: for the high-resolution FERET and CMU PIE face libraries, the resolution is respectively as follows: the resolution of the FERET face library is 72 multiplied by 72, the resolution of the CMU PIE face library is 32 multiplied by 28, the FERET face library is multiplied by 4 and 9, the CMU PIE face library is subjected to 2 and 4 times of smooth down-sampling, the resolution of the FERET face library in the generated LR training set is 18 multiplied by 18 and 8 multiplied by 8, and the resolution of the CMU PIE face library is 16 multiplied by 14 and 8 multiplied by 7.
The step 2 specifically comprises the following steps:
step 2.1, based on the coupling mapping learning method, simultaneously mapping the face images in the HR training set and the LR training set to a public feature subspace, and expressing that:
wherein the HR feature vector is:
corresponding LR feature vectors:
f
Hmapping function representing HR face image to common feature subspace, corresponding f
LRepresenting a mapping function of the LR face image to a public feature subspace, and d representing the dimension of the public feature subspace;
step 2.2, set f
H(x)=P
H Tx and f
L(x)=P
L Tx, the formula of step 2.1 is expressed in a matrixing way, and is expressed as:
wherein P is
HMapping matrices and P for HR coupling
LThe mapping matrix is coupled to the LR.
The step 3 specifically comprises the following steps:
step 3.1, adding the local geometric structure information and the discrimination information of the sample into the formula subjected to matrixing in the step 2, and expressing as follows:
wherein a and b respectively represent the weight factors of the discrimination items in the class and the discrimination items between the classes,
is a matrix of local similarities which is a matrix of local similarities,
representing an intra-class discrimination matrix;
represents an inter-class discrimination matrix, i is 1,2, 3 … N
t,j=1、2、 3…N
t;
Step 3.2, setting
And
equation (1) is then expressed as:
wherein D is
s、D
wAnd D
bAre diagonal matrices, respectively defined as
And
step 3.3, setting
And
equation (2) can be simplified to J (P)
H,P
L)=tr(P
TYGY
TP);
Step 3.4, the objective function of the formula (2) is minimized to solve the following optimization problem: j (P)H,PL)s.t.PTYYTP=Iand PTY1 ═ 0, where I is a unit array of size d × d, 1 ═ 1,1]TIs composed of 2N t1 vector of term, set
Andthe solution P of the optimization problem is obtained by solving the 2 nd to (d +1) th generalized eigenvectors Ep ═ λ Fp (3) of P,
expanding equation (3) yields:
the formula (4) is simplified to obtain:
two coupling mapping matrixes P can be obtained by jointly solving the formula (5)HAnd PL。
Local similarity matrix
Is represented as follows:
wherein n is the number of neighborhood samples belonging to the HR training face image of the same class;
represents a gaussian kernel width; c is a scale factor;
in-class discriminant matrix
Is represented as follows:
wherein k iswRepresenting the number of neighborhood samples in the HR training face image class;
inter-class discrimination matrix
Is represented as follows:
wherein k isbAnd the number of neighborhood samples among the HR training face image classes is represented.
The step 4 specifically comprises the following steps:
and respectively carrying out smooth downsampling on the other half of high-resolution face images in the HR image set to generate an LR test set:
wherein
Representing the ith low-resolution test face image,
representing the total number of the images of the test set;
wherein, the high resolution means that the resolution of the FERET face library is 72 multiplied by 72, and the resolution of the CMU PIE face library is 32 multiplied by 28;
respectively performing smooth downsampling on the samples: 4 times and 9 times of FERET face library, and 2 times and 4 times of CMU PIE face library;
low resolution means respectively: the FERET face library resolution is 18 × 18 and 8 × 8, and the CMU PIE face library resolution is 16 × 14 and 8 × 7.
The step 5 specifically comprises the following steps:
and transforming the HR image set and the LR test set to a public feature subspace to obtain HR-LR face mapping features, which specifically comprises the following steps:
two coupling matrixes P solved according to the training stage
HAnd P
LRespectively collecting HR images X
GAnd LR test set X
PTransforming to common feature subspace to obtain HR-LR face mapping features
And
the
step 6 specifically comprises the following steps: applying nearest neighbor classifier to HR-LR face mapping features projected in common feature subspace
Classifying to obtain the face mapping characteristics
Class label of
The invention has the beneficial effects that:
(1) according to the method, the local geometric structure information of the sample is added into the target function, so that the local neighborhood relationship of the HR-LR face image in the respective original feature space can be effectively reserved, and the separability of the sample in the projected public feature subspace is greatly enhanced;
(2) the method simultaneously considers the intra-class and inter-class discrimination information of the sample, and effectively improves the discrimination of the sample in the projected public feature subspace;
(3) the invention fully considers the discrimination information and the label information of the sample, and utilizes the local manifold geometric structure information and the label information of the HR-LR face image to ensure that the distances of the HR-LR face images of the same class in the obtained public characteristic subspace are as close as possible, and the distances of the HR-LR face images of different classes in the public characteristic subspace are as distant as possible, thereby enhancing the discrimination capability and the separability of the coupling mapping relation matrix and improving the identification performance of the LR face.
Detailed Description
The present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.
The invention relates to a low-resolution face recognition method based on multi-manifold coupling mapping, which comprises a training stage and a testing stage, and is implemented according to the following steps:
first, training phase
Step 1, selecting N HR face images from a standard face database (FERET and CMU PIE face database) to form an HR image set
Randomly selecting half of the face images containing each person from the HR image set as an HR training set:
performing smooth down-sampling on HR training sets with high resolution (the resolution of a FERET face library is 72 multiplied by 72, and the resolution of a CMU PIE face library is 32 multiplied by 28) (4 times and 9 times of the FERET face library, and 2 times and 4 times of the CMU PIE face library) respectively to generate low resolution (the resolution of the FERET face library is 18 multiplied by 18 and 8 multiplied by 8, and the resolution of the CMU PIE face library is 16 multiplied by 14 and 8 multiplied by 7) training sets:
wherein
Representing the ith low-resolution face image,
representing the ith high-resolution face image,
representing the total number of the images of the training set; the class labels of the training face image samples are as follows:
as shown in fig. 1, step 2, based on the coupling mapping learning method, simultaneously mapping the face images in the HR training set and the LR training set to the common feature subspace, obtaining a formula based on the coupling mapping and performing matrixing, specifically:
step 2.1, based on the coupling mapping learning method, simultaneously mapping the face images in the HR training set and the LR training set to a public feature subspace, and expressing that:
wherein the HR feature vector is:
corresponding LR feature vectors:
f
Hmapping function representing HR face image to common feature subspace, corresponding f
LRepresenting a mapping function of the LR face image to a public feature subspace, and d representing the dimension of the public feature subspace;
step 2.2, set f
H(x)=P
H Tx and f
L(x)=P
L Tx, the formula of step 2.1 is expressed in a matrixing way, and is expressed as:
wherein P is
HMapping matrices and P for HR coupling
LMapping a matrix for LR coupling;
step 3, adding the local geometric structure information and the discrimination information of the sample into the formula subjected to matrixing in the step 2, and solving an HR coupling mapping matrix PHAnd LR coupling mapping matrix PLAnd setting a dieType parameters a, b, c, n, kwAnd kbThe method specifically comprises the following steps:
step 3.1, adding the local geometric structure information and the discrimination information of the sample into the formula subjected to matrixing in the step 2, and expressing as follows:
wherein a and b respectively represent the weight factors of the discrimination items in the class and the discrimination items between the classes,
is a matrix of local similarities which is a matrix of local similarities,
representing an intra-class discrimination matrix;
represents an inter-class discrimination matrix, i is 1,2, 3 … N
t,j=1、2、 3…N
t;
Step 3.2, setting
And
equation (1) is then expressed as:
wherein D is
s、D
wAnd D
bAre diagonal matrices, respectively defined as
Step 3.3, setting
And
equation (2) can be simplified to J (P)
H,P
L)=tr(P
TYGY
TP);
Step 3.4, the objective function of the formula (2) is minimized to solve the following optimization problem: j (P)H,PL)s.t.PTYYTP=Iand PTY1 ═ 0, where I is a unit array of size d × d, 1 ═ 1,1]TIs composed of 2N t1 vector of term, set
And
the solution P to the optimization problem can be obtained by solving the 2 nd to (d +1) th generalized eigenvectors Ep ═ λ Fp (3) of P,
expanding equation (3) yields:
the formula (4) is simplified to obtain:
two coupling mapping matrixes P can be obtained by jointly solving the formula (5)HAnd PL;
Local similarity matrixIs represented as follows:
wherein n is the number of neighborhood samples belonging to the HR training face image of the same class;
represents the Gaussian kernel width; c is a scale factor;
in-class discriminant matrix
Is represented as follows:
wherein k iswRepresenting the number of neighborhood samples in the HR training face image class;
inter-class discrimination matrix
Is represented as follows:
wherein k isbAnd the number of neighborhood samples among the HR training face image classes is represented.
Second, testing stage
And 4, respectively carrying out smooth downsampling on the other half of high-resolution face images in the HR image set to generate an LR test set:
wherein
Representing the ith low-resolution test face image,
representing the total number of the images of the test set;
wherein, the high resolution means that the resolution of the FERET face library is 72 multiplied by 72, and the resolution of the CMU PIE face library is 32 multiplied by 28;
respectively performing smooth downsampling on the samples: 4 times and 9 times of FERET face library, and 2 times and 4 times of CMU PIE face library;
low resolution means respectively: the resolution of the FERET face library is 18 multiplied by 18 and 8 multiplied by 8, and the resolution of the CMU PIE face library is 16 multiplied by 14 and 8 multiplied by 7;
and 5, transforming the HR image set and the LR test set to a public feature subspace to obtain an HR-LR face mapping feature
And
the method specifically comprises the following steps:
two coupling matrixes P solved according to the training stage
HAnd P
LRespectively collecting HR images X
GAnd LR test set X
PTransforming to common feature subspace to obtain HR-LR face mapping features
And
step 6, applying the nearest neighbor classifier to the HR-LR face mapping feature projected in the public feature subspace
Classifying to obtain the face mapping characteristics
Class label of
To verify the effectiveness of the present invention, the following simulations were performed:
on the same training set and test set, some benchmark methods for extracting face features by using Principal Component Analysis (PCA) and Linear Discriminant Analysis (LDA) are selected in the form of comparison experiments, such as HR-PCA (extracting features directly using PCA for HR face images), HR-LDA (extracting features directly using LDA for HR face images), customer-PCA (performing customer interpolation for LR face images and extracting features using PCA), customer-LDA (performing customer interpolation for LR face images and extracting features using PCA), and a Coupled Local Preserving Mapping (CLPMs) method for comparison to verify the effectiveness of the present invention, wherein the CLPMs method specifically refers to LI B, IEEE change, scan G, low-preserving method [ process ] for example, 2010,17(1):20-23.".
First, the invention uses Rank-1 and 8 × 8(FERET face library) and 8 × 7(CMU PIE face library) to carry out experiments, and analyzes the influence of characteristic dimension on the recognition effect. As can be seen from the simulation results of fig. 2 and 3: the recognition effect of the invention on 2 standard face libraries approaches or exceeds the HR-LDA reference method. The identification performance of the method is far superior to that of other methods, and the identification performance is distributed in a dimension segment with lower feature dimension. The method not only considers the local geometric structure information of the sample, but also considers the intra-class and inter-class discrimination information of the sample, so that the dual mapping obtained by learning can effectively improve the discrimination and the separability of the sample.
Experiment two, Rank-n is an important index for evaluating the performance of the recognition algorithm in pattern recognition, and is used for calculating the probability that the first n face images in the matching result contain correct matching. After the matched faces are sorted from high to low in the candidate set according to the similarity, the more the correctly matched faces are sorted, the better the algorithm effect is. In the experiment, Rank-n is adopted to evaluate the performance of the invention, and figures 4 and 5 show the identification performance of the invention under different Rank levels. As can be seen from the simulation results of fig. 4: taking Rank-1 of a FERET face library as an example, the probability of matching a target face for the first time in n (n is 1,2, …,10) most similar faces reaches about 94 percent, and on 2 standard face libraries, the highest recognition rate of the method is obviously superior to other methods at different Rank levels, and the recognition rate slowly rises along with the gradual increase of the Rank levels and finally tends to be flat. The experiment fully shows that the invention has better stability.
And thirdly, two resolutions are respectively set for each standard face library in the experiment to evaluate the recognition performance of the invention and analyze the influence of the resolutions on the recognition effect, wherein the resolutions of the FERET face library are respectively 8 × 8 and 18 × 18, and the resolutions of the CMU PIE face library are respectively 8 × 7 and 16 × 14. As can be seen from the simulation results of fig. 6 and 7: the recognition effect of the invention on 2 standard face libraries is better than that of other methods, and the recognition effect is not worse than that of other methods due to the influence of resolution, which fully shows that the invention has good robustness on the resolution of the face image.
The results of the three experiments show that compared with the existing low-resolution face recognition method based on coupling mapping, the method has stronger discrimination and separability on the sample, and the recognition performance is far better than that of other similar methods.