CN107451537A - Face identification method based on deep learning multilayer Non-negative Matrix Factorization - Google Patents

Abstract

The invention discloses a kind of face identification method based on deep learning multilayer Non-negative Matrix Factorization, mainly solves the problems, such as that existing face recognition technology discrimination under complicated cosmetic variation is low.Its technical scheme is：1. utilize the characteristic of VGG Face extraction training samples and each channel data of test sample；2. the characteristic of each channel data of pair training sample repeats the characteristic extraction procedure of L normalization, nonlinear transformation and matrix decomposition respectively, low-rank robust features are obtained；3. build K nearest neighbor classifier；4. the characteristic of each channel data of test sample is projected respectively, projection coefficient vector is obtained；5. projection coefficient vector is input into K nearest neighbor classifier to be classified；6. integrating the classification results of K nearest neighbor classifier, the recognition result of test sample is obtained.The present invention improves the face identification rate under complicated cosmetic variation, can be applied to identity authentication and information security field.

Description

Face recognition method based on deep learning multi-layer non-negative matrix factorization

technical field

The invention belongs to the technical field of image processing, in particular to a face image recognition method, which can be applied to the fields of identification and information security.

Background technique

With the continuous development of human society, face recognition has been widely used in security, finance, e-government and many other fields. Improving the performance of face recognition is conducive to expanding the application of face recognition. The current main research on face recognition is to extract effective, robust and more discriminative features and design classifiers with better classification ability. Selecting more robust and discriminative features and designing a classifier with good classification ability are the keys to improve the robustness of face recognition.

Non-negative matrix decomposition is a feature extraction method for matrix decomposition under non-negative constraints. It has good data representation ability and can greatly reduce the dimensionality of data features, and its decomposition characteristics are in line with the intuitive experience of human visual perception. Decomposition The results have interpretable and clear physical meaning. The basic non-negative matrix factorization NMF directly decomposes the original coefficient matrix into a base matrix and a coefficient matrix, and requires both the base matrix and the coefficient matrix to be non-negative, which shows that there is only an additive combination in the non-negative matrix factorization NMF. Therefore, non-negative matrix factorization NMF can be regarded as a model based on partial representation, which can provide the local structure of the observed data, but in some cases, the NMF algorithm will also give global features, resulting in limited classification performance.

Deep learning is a new research direction of feature representation in the field of machine learning. In recent years, breakthroughs have been made in various applications such as speech recognition and computer vision. Deep learning forms more abstract high-level representations or features by combining low-level features. In the deep learning model, there are more nonlinear transformation layers and stronger generalization ability. But in practical applications, appearance changes caused by factors such as head posture, lighting, occlusion, etc. will lead to a decline in the performance of deep learning, and so far there is no good solution.

Contents of the invention

In view of the disadvantages of the prior art described above, the purpose of the present invention is to provide a face recognition method based on deep learning multi-layer non-negative matrix factorization, to obtain low-rank robust features that are more discriminative in depth, and improve complex Face recognition rate under appearance variation.

The technical key to realize the present invention is that on the basis of deep learning, a new multi-layer non-negative matrix decomposition is introduced to improve the existing deep learning methods. Specifically, the present invention obtains a more discriminative low-rank feature representation by performing multiple non-negative matrix decompositions on the sample features obtained by deep learning, thereby improving the face recognition rate. The steps include the following:

(1) Input the data of each channel of the training sample into the VGG-Face deep convolutional neural network to obtain the characteristic data X(k) of each channel data of the training sample, where k=1,2,..., K, K is the channel number of training samples;

(2) Carry out the feature extraction process of normalization, nonlinear transformation and matrix decomposition to the characteristic data X (k) that step (1) obtains respectively, obtain coefficient matrix H (k);

(3) Repeat the feature extraction process in step (2) L times to obtain low-rank robust features h _j (k), where j=1,2,...,n, n is the total number of training samples;

(4) Construct K nearest neighbor classifiers according to the low-rank robust features h _j (k) obtained in step (3);

(5) Input each channel data of the test sample into the VGG-Face deep convolutional neural network to obtain the characteristic data Y(k) of each channel data of the test sample;

(6) Perform the projection process according to the characteristic data Y(k) obtained in step (5), and obtain the projection coefficient vector

(7) The projection coefficient vector obtained in step (6) Input to the K nearest neighbor classifiers to obtain the classification result of each channel of the test sample, where i=1,2,...,e, e is the total number of test samples;

(8) Combining the classification results of each channel of the test sample obtained in step (7) to obtain the classification result of the test sample.

Compared with the prior art, the present invention has the following advantages:

1) The present invention combines multi-layer non-negative matrix decomposition on the basis of deep learning, which can obtain more discriminative feature representation;

2) The present invention further improves the face recognition rate under complex appearance changes by synthesizing the classification results of different channels.

Description of drawings

Fig. 1 is the realization flowchart of the present invention.

detailed description

With reference to Fig. 1, the face recognition step of the present invention based on deep learning multi-layer non-negative matrix decomposition is as follows:

Step 1, obtain the feature data X(k) of each channel data of the training sample.

(1a) Obtain the face data set V ^train as a training data set, the total number of training samples in the training data set is n, the number of categories in the training data set is c, and each training sample in the training data set is equally divided into K Area, each area is used as one channel data of the training sample, and the training sample contains K channel data in total;

(1b) According to the training data set, under the Linux operating system, use the Caffe deep learning framework to fine-tune the parameters of the VGG-Face deep convolutional neural network;

(1c) Input each channel data of each training sample in the training data set into the VGG-Face deep convolutional neural network to obtain the feature data X(k) of training each channel data, where k=1,2, ...,K; K is the channel number of training samples.

Step 2, according to the feature data X(k), obtain the coefficient matrix H(k).

Perform the feature extraction process of normalization, nonlinear transformation and matrix decomposition on the feature data X(k) to obtain the coefficient matrix H(k);

(2a) Normalize the feature data X(k) using the L2 norm;

(2b) use the sigmoid function to carry out non-linear transformation to the result after step (2a) normalization processing, obtain the transformed result B(k);

(2c) Use soft-constrained non-negative matrix decomposition to perform matrix decomposition on the result B(k) after nonlinear transformation in step (2b), and obtain B(k)≈Z(k)A(k)F(k), where , B(k) is a matrix of order m×n, Z(k) is a base matrix of order m×φ, A(k) is an auxiliary matrix of order φ×c, and F(k) is a predicted label of order c×n Matrix, m is the original feature dimension, φ is the decomposition dimension, c is the number of categories, and n is the total number of training samples;

(2c1) Randomly initialize the base matrix Z ⁽¹⁾ (k), auxiliary matrix A ⁽¹⁾ (k) and predictive label matrix F ⁽¹⁾ (k) as the result of one iteration, where the base matrix Z ^{(1 )} Any element in (k) satisfies is the pth row q column element of the base matrix Z ⁽¹⁾ (k); any element in the auxiliary matrix A ⁽¹⁾ (k) satisfies is the element in row β and column α of the auxiliary matrix A ⁽¹⁾ (k); any element in the predicted label matrix F ⁽¹⁾ (k) satisfies is the γth row of the predicted label matrix F ⁽¹⁾ (k) Column elements; p=1,2,...,m, q=1,2,...,φ, α=1,2,...,φ, β=1,2,...,c , γ=1,2,...,c,

(2c2) Update the elements Z _{p, q} in the base matrix Z according to the following formula:

Among them, t is the number of iterations, t=2,..., iter, iter is the maximum number of iterations, T is the matrix transpose, is the pth row q column element of the unnormalized base matrix Z ^(t)' (k) obtained after iteration t times;

(2c3) normalize the base matrix Z ^(t) '(k) obtained in step (2c2), obtain the base matrix Z ( ^t ) (k) of iteration t times;

(2c4) Update the elements A _{α, β} (k) in the auxiliary matrix A(k) according to the following formula:

in, For the auxiliary matrix A ^(t) (k) obtained after the iteration t times, the α row and β column element; A ^(t) (k) is the auxiliary matrix obtained after the iteration t times;

(2c5) According to the following formula, predict the elements in the label matrix F(k) Make an update:

in, is the γ-th row of the predicted label matrix F ^(t) (k) after iteration t times Column elements; F ^(t) (k) is the predicted label matrix after iteration t times; λ is the coefficient of the regular term; is the γ-th row of the pre-defined local label matrix C(k) column element;

(2c6) Determine whether the number of iterations t has reached the maximum number of iterations iter: if so, stop the iteration, and use the base matrix Z ^(iter) (k), auxiliary matrix A ^(iter) (k) and predicted label obtained in the iter iteration Matrix F ^(iter) (k), as the final base matrix Z(k), auxiliary matrix A(k) and predicted label matrix F(k); otherwise, return to step (2c2);

(2d) According to the auxiliary matrix A(k) obtained after the soft-constrained non-negative matrix decomposition in step (2c) and the predicted label matrix F(k), the coefficient matrix is obtained: H(k)=A(k)F(k).

Step 3, obtain the low-rank robust features h(k) of the training samples.

Repeat the feature extraction process in step 2 to obtain the low-rank robust feature h(k) of each channel feature data X(k) of the training sample;

(3a) Process the feature data X(k) of each channel of the training sample according to step 2 to obtain the first-layer base matrix Z ¹ (k) and the first-layer coefficient matrix H ¹ (k);

(3b) Process the coefficient matrix H ¹ (k) of the first layer obtained in step (3a) according to step 2 to obtain the base matrix Z ² (k) of the second layer and the coefficient matrix H ² (k) of the second layer;

(3c) Continue to repeat the same steps according to steps (3a) and (3b), and obtain the base matrix Z ^l (k) of the l-th layer and the coefficient matrix of the l-th layer according to the coefficient matrix H ^l-1 (k) of the l-1th layer H ^l (k), until the number of repetitions l=L, the base matrix Z ^L (k) of the L layer and the coefficient matrix H ^L (k) of the L layer are obtained, wherein, l=2,...,L, L is The number of layers of multi-level non-negative matrix factorization;

(3d) According to the L-th layer coefficient matrix H ^L (k) obtained in step (3c), obtain the low-rank robust features h _j (k) of each channel of the training sample, where j=1,2,..., n.

Step 4, according to the low-rank robust feature h _j (k) obtained in step 3, construct K nearest neighbor classifiers.

(4a) From the results obtained in step 3, select the low-rank robust features h _j (k) of the kth channel of each training sample to form a feature set;

(4b) form a nearest neighbor classifier according to the feature set obtained in step (4a);

(4c) Repeat steps (4a) and (4b) for different channels to obtain K nearest neighbor classifiers.

Step 5, obtain the characteristic data Y(k) of each channel data of the test sample.

(5a) Obtain the face data set V ^test with the same attributes as the training data set as the test data set, the total number of test samples in the test data set is e, the number of categories in the test data set is c, and each The test sample is divided into K channel data according to step (1a);

(5b) according to step (1b), the parameter of VGG-Face deep convolutional neural network is set;

(5c) Input the data of each channel of the test sample into the VGG-Face deep convolutional neural network to obtain the feature data Y(k) of each channel data of the test sample.

Step 6, respectively project the characteristic data Y(k) of each channel data of the test sample obtained in step 5, and output the projection coefficient vector

(6a) Perform normalization processing, nonlinear transformation and projection transformation on the characteristic data Y(k) of the test sample to obtain the projection matrix of the first layer

(6a1) Normalize the feature data Y(k) of the test sample using the L2 norm;

(6a2) Use the Sigmoid function to perform nonlinear transformation on the result obtained after the normalization process in step (6a1), and obtain the transformation result f(Y(k)) after the nonlinear transformation, where f(·) means using Sigmoid function for nonlinear transformation;

(6a3) Perform projection transformation on the first-layer base matrix Z ¹ (k) obtained in step (3a) to obtain the first-layer projection matrix : in, Indicates the generalized inverse operation;

(6b) The first layer projection matrix obtained according to step (6a) Perform the same process as the base matrix Z ² (k) of the second layer to obtain the projection matrix of the second layer

(6c) Continue to repeat the same steps according to steps (6a) and (6b), according to the l-1th layer projection matrix and the base matrix Z ^l (k) of the l-th layer to obtain the projection matrix of the l-th layer Until the number of repetitions l=L, the L-th layer projection matrix is obtained Among them, l=2,...,L;

(6d) The L-th layer projection matrix obtained according to step (6c) Get the projection coefficient vector for each test sample Wherein, i=1, 2, . . . , e.

Step 7, the projection coefficient vector obtained in step 6 Input to the K nearest neighbor classifiers to get the classification results of each channel of the test sample.

(7a) Calculate the low-rank robust feature h _j (k) of the training sample and the projection coefficient vector of the test sample The low-dimensional Euclidean distance between get distance set Among them, j=1,2,...,n, i∈{1,2,...,e}, ||·|| ₂ means 2-norm;

(7b) The distance set obtained according to step (7a) The minimum value in the distance set The category of the corresponding ξ-th training sample is used as the classification result of the i-th test sample on the k-th nearest neighbor classifier, where ξ∈{1,2,...,n};

(7c) Classify the K channels of each test sample according to steps (7a) and (7b), and obtain the classification results of each test sample on K nearest neighbor classifiers.

In step 8, the classification results of each channel of the test sample obtained in step 7 are integrated to obtain the final classification result of the test sample.

(8a) According to the classification results of each test sample on the K nearest neighbor classifiers obtained in step 7, respectively count the number of correctly classified test samples CN _k on each nearest neighbor classifier, and calculate each nearest neighbor classification The recognition rate of the device:

Among them, CN _k is the number of correctly classified test samples on the kth nearest neighbor classifier, and o _k is the recognition rate of the kth nearest neighbor classifier;

(8b) Calculate the linear weight coefficients α _k of the K nearest neighbor classifiers respectively according to the recognition rates of the K nearest neighbor classifiers obtained in step (8a):

(8c) According to the linear weight coefficient α _k obtained in step (8b), calculate the test sample K channel projection coefficient vectors The weighted distance between the K channel low-rank robust features h _j (k) of the training sample:

, get the weighted distance set {d _1i ,d _2i ,...,d _ji ,...,d _ni };

(8d) According to the weighted distance set {d _1i ,d _2i ,...,d _ji ,...,d _ni } obtained in step (8c), train the ωth corresponding to the minimum value d _ωi in the weighted distance set The category of the sample is the classification result of the test sample, where ω∈{1,2,...,n}.

The above description is only a specific example of the present invention, and does not constitute any limitation to the present invention. Obviously, for those skilled in the art, after understanding the content and principles of the present invention, it is possible without departing from the principles and structures of the present invention. In some cases, various modifications and changes in form and details are made, but these modifications and changes based on the idea of the present invention are still within the protection scope of the claims of the present invention.

Claims (7)

1. A face recognition method based on deep learning multi-layer non-negative matrix factorization, including: (1) Input the data of each channel of the training sample into the VGG-Face deep convolutional neural network to obtain the characteristic data X(k) of each channel data of the training sample, where k=1,2,..., K, K is the channel number of training samples; (2) Carry out the feature extraction process of normalization, nonlinear transformation and matrix decomposition to the characteristic data X (k) that step (1) obtains respectively, obtain coefficient matrix H (k); (3) Repeat the feature extraction process in step (2) L times to obtain low-rank robust features h _j (k), where j=1,2,...,n, n is the total number of training samples; (4) Construct K nearest neighbor classifiers according to the low-rank robust features h _j (k) obtained in step (3); (5) Input each channel data of the test sample into the VGG-Face deep convolutional neural network to obtain the characteristic data Y(k) of each channel data of the test sample; (6) Perform the projection process according to the characteristic data Y(k) obtained in step (5), and obtain the projection coefficient vector (7) The projection coefficient vector obtained in step (6) Input to the K nearest neighbor classifiers to obtain the classification result of each channel of the test sample, where i=1,2,...,e, e is the total number of test samples; (8) Combining the classification results of each channel of the test sample obtained in step (7) to obtain the classification result of the test sample. 2. The method according to claim 1, wherein the implementation steps of the step (2) are as follows: (2a) Normalize the feature data X(k) using the L2 norm; (2b) use the sigmoid function to carry out non-linear transformation to the result after step (2a) normalization processing, obtain the transformed result B(k); (2c) Use soft-constrained non-negative matrix decomposition to perform matrix decomposition on the result B(k) after nonlinear transformation in step (2b), and obtain B(k)≈Z(k)A(k)F(k), where , B(k) is a matrix of order m×n, Z(k) is a base matrix of order m×φ, A(k) is an auxiliary matrix of order φ×c, and F(k) is a predicted label of order c×n Matrix, m is the original feature dimension, φ is the decomposition dimension, c is the number of categories, and n is the total number of training samples; (2d) According to the auxiliary matrix A(k) obtained after the soft-constrained non-negative matrix decomposition in step (2c) and the predicted label matrix F(k), the coefficient matrix is obtained: H(k)=A(k)F(k). 3. The method according to claim 2, wherein in the step (2c), use soft constraint non-negative matrix decomposition to carry out matrix decomposition to the result B (k) after nonlinear transformation in the step (2b), carry out as follows: (2c1) Randomly initialize the base matrix Z ⁽¹⁾ (k), auxiliary matrix A ⁽¹⁾ (k) and predictive label matrix F ⁽¹⁾ (k) as the result of one iteration, where the base matrix Z ^{(1 )} Any element in (k) satisfies is the pth row q column element of the base matrix Z ⁽¹⁾ (k); any element in the auxiliary matrix A ⁽¹⁾ (k) satisfies is the element in row β and column α of the auxiliary matrix A ⁽¹⁾ (k); any element in the predicted label matrix F ⁽¹⁾ (k) satisfies is the γth row of the predicted label matrix F ⁽¹⁾ (k) Column elements; p=1,2,...,m, q=1,2,...,φ, α=1,2,...,φ, β=1,2,...,c , γ=1,2,...,c, (2c2) Update the elements Z _{p, q} in the base matrix Z according to the following formula: <mrow><msubsup><mi>Z</mi><mrow><mi>p</mi><mo>,</mo><mi>q</mi></mrow><mrow><mo>(</mo><mi>t</mi><mo>)</mo><mo>&amp;prime;</mo></mrow></msubsup><mrow><mo>(</mo><mi>k</mi><mo>)</mo></mrow><mo>=</mo><msubsup><mi>Z</mi><mrow><mi>p</mi><mo>,</mo><mi>q</mi></mrow><mrow><mo>(</mo><mi>t</mi><mo>-</mo><mn>1</mn><mo>)</mo></mrow></msubsup><mrow><mo>(</mo><mi>k</mi><mo>)</mo></mrow><mfrac><msubsup><mrow><mo>(</mo><mi>B</mi><mo>(</mo><mi>k</mi><mo>)</mo><msup><mi>F</mi><mi>T</mi></msup><mo>(</mo><mi>k</mi><mo>)</mo><msup><mi>A</mi><mi>T</mi></msup><mo>(</mo><mi>k</mi><mo>)</mo><mo>)</mo></mrow><mrow><mi>p</mi><mo>,</mo><mi>q</mi></mrow><mrow><mo>(</mo><mi>t</mi><mo>-</mo><mn>1</mn><mo>)</mo></mrow></msubsup><msubsup><mrow><mo>(</mo><mi>Z</mi><mo>(</mo><mi>k</mi><mo>)</mo><mi>A</mi><mo>(</mo><mi>k</mi><mo>)</mo><mi>F</mi><mo>(</mo><mi>k</mi><mo>)</mo><msup><mi>F</mi><mi>T</mi></msup><mo>(</mo><mi>k</mi><mo>)</mo><msup><mi>A</mi><mi>T</mi></msup><mo>(</mo><mi>k</mi><mo>)</mo><mo>)</mo></mrow><mrow><mi>p</mi><mo>,</mo><mi>q</mi></mrow><mrow><mo>(</mo><mi>t</mi><mo>-</mo><mn>1</mn><mo>)</mo></mrow></msubsup></mfrac><mo>,</mo></mrow> Among them, t is the number of iterations, t=2,..., iter, iter is the maximum number of iterations, T is the matrix transpose, It is the pth row q column element of the unnormalized base matrix Z ^(t) '(k) obtained after iteration t times; (2c3) normalize the base matrix Z ^(t) '(k) obtained in step (2c2), obtain the base matrix Z ( ^t ) (k) of iteration t times; (2c4) Update the elements A _{α, β} (k) in the auxiliary matrix A(k) according to the following formula: <mrow><msubsup><mi>A</mi><mrow><mi>&amp;alpha;</mi><mo>,</mo><mi>&amp;beta;</mi></mrow><mrow><mo>(</mo><mi>t</mi><mo>)</mo></mrow></msubsup><mrow><mo>(</mo><mi>k</mi><mo>)</mo></mrow><mo>=</mo><msubsup><mi>A</mi><mrow><mi>&amp;alpha;</mi><mo>,</mo><mi>&amp;beta;</mi></mrow><mrow><mo>(</mo><mi>t</mi><mo>-</mo><mn>1</mn><mo>)</mo></mrow></msubsup><mrow><mo>(</mo><mi>k</mi><mo>)</mo></mrow><mfrac><msubsup><mrow><mo>(</mo><msup><mi>Z</mi><mi>T</mi></msup><mo>(</mo><mi>k</mi><mo>)</mo><mi>B</mi><mo>(</mo><mi>k</mi><mo>)</mo><msup><mi>F</mi><mi>T</mi></msup><mo>(</mo><mi>k</mi><mo>)</mo><mo>)</mo></mrow><mrow><mi>&amp;alpha;</mi><mo>,</mo><mi>&amp;beta;</mi></mrow><mrow><mo>(</mo><mi>t</mi><mo>-</mo><mn>1</mn><mo>)</mo></mrow></msubsup><msubsup><mrow><mo>(</mo><msup><mi>Z</mi><mi>T</mi></msup><mo>(</mo><mi>k</mi><mo>)</mo><mi>Z</mi><mo>(</mo><mi>k</mi><mo>)</mo><mi>A</mi><mo>(</mo><mi>k</mi><mo>)</mo><mi>F</mi><mo>(</mo><mi>k</mi><mo>)</mo><msup><mi>F</mi><mi>T</mi></msup><mo>(</mo><mi>k</mi><mo>)</mo><mo>)</mo></mrow><mrow><mi>&amp;alpha;</mi><mo>,</mo><mi>&amp;beta;</mi></mrow><mrow><mo>(</mo><mi>t</mi><mo>-</mo><mn>1</mn><mo>)</mo></mrow></msubsup></mfrac><mo>,</mo></mrow> in, For the auxiliary matrix A ^(t) (k) obtained after the iteration t times, the α row and β column element; A ^(t) (k) is the auxiliary matrix obtained after the iteration t times; (2c5) According to the following formula, predict the elements in the label matrix F(k) Make an update: in, is the γ-th row of the predicted label matrix F ^(t) (k) after iteration t times Column elements; F ^(t) (k) is the predicted label matrix after iteration t times; λ is the coefficient of the regular term; is the γ-th row of the pre-defined local label matrix C(k) column element; (2c6) Determine whether the number of iterations t has reached the maximum number of iterations iter: if so, stop the iteration, and use the base matrix Z ^(iter) (k), auxiliary matrix A ^(iter) (k) and predicted label obtained in the iter iteration Matrix F ^(iter) (k), as the final base matrix Z(k), auxiliary matrix A(k) and predicted label matrix F(k); otherwise, return to step (2c2). 4. The method according to claim 1, wherein the implementation steps of the step (3) are as follows: (3a) Process the feature data X(k) of each channel of the training sample according to step (2), and obtain the first-layer base matrix Z ¹ (k) and the first-layer coefficient matrix H ¹ (k), where k =1,2,...,K; (3b) Process the first layer coefficient matrix H ¹ (k) obtained in step (3a) according to step (2), and obtain the second layer basic matrix Z ² (k) and the second layer coefficient matrix H ² (k) ; (3c) Continue to repeat the same steps according to steps (3a) and (3b), and obtain the base matrix Z ^l (k) of the l-th layer and the coefficient matrix of the l-th layer according to the coefficient matrix H ^l-1 (k) of the l-1th layer H ^l (k), until the number of repetitions l=L, the base matrix Z ^L (k) of the L layer and the coefficient matrix H ^L (k) of the L layer are obtained, wherein, l=2,...,L, L is The number of layers of multi-level non-negative matrix factorization; (3d) According to the L-th layer coefficient matrix H ^L (k) obtained in step (3c), obtain the low-rank robust features h _j (k) of each channel of the training sample, where j=1,2,..., n. 5. The method according to claim 1, wherein the implementation steps of the step (6) are as follows: (6a) Perform normalization processing, nonlinear transformation and projection transformation process on the characteristic data Y(k) of the test sample to obtain the projection matrix Among them, k∈{1,2,...,K}, K is the number of sample channels; (6a1) Normalize the feature data Y(k) of the test sample using the L2 norm; (6a2) Use the Sigmoid function to perform nonlinear transformation on the result obtained after the normalization process in step (6a1), and obtain the transformation result f(Y(k)) after the nonlinear transformation, where f(·) means using Sigmoid function for nonlinear transformation; (6a3) Perform projection transformation on the first-layer base matrix Z ¹ (k) obtained in step (3a) to obtain the first-layer projection matrix : in, Indicates the generalized inverse operation; (6b) The first layer projection matrix obtained according to step (6a) Perform the same process as the base matrix Z ² (k) of the second layer to obtain the projection matrix of the second layer (6c) Continue to repeat the same steps according to steps (6a) and (6b), according to the l-1th layer projection matrix and the base matrix Z ^l (k) of the l-th layer to obtain the projection matrix of the l-th layer Until the number of repetitions l=L, the L-th layer projection matrix is obtained Wherein, l=2,...,L, L is the number of layers of multi-layer non-negative matrix decomposition; (6d) The L-th layer projection matrix obtained according to step (6c) Get the projection coefficient vector for each test sample Wherein, i=1, 2, . . . , e. 6. The method according to claim 1, wherein said step (7) is carried out as follows: (7a) Calculate the low-rank robust feature h _j (k) of the training sample and the projection coefficient vector of the test sample The low-dimensional Euclidean distance between get distance set Among them, j=1,2,...,n, k∈{1,2,...,K}, i∈{1,2,...,e}, ||·|| ₂ means 2 norm; (7b) The distance set obtained according to step (7a) The minimum value in the distance set The category of the corresponding ξ-th training sample is used as the classification result of the i-th test sample on the k-th nearest neighbor classifier, where ξ∈{1,2,...,n}; (7c) Classify the K channels of each test sample according to steps (7a) and (7b), and obtain the classification results of each test sample on K nearest neighbor classifiers. 7. The method according to claim 1, wherein said step (8) is carried out as follows: (8a) According to the classification results of each test sample obtained in step (7) on the K nearest neighbor classifiers, respectively count the number of correctly classified test samples CN _k on each nearest neighbor classifier, and calculate each nearest neighbor classifier. The recognition rate of the neighbor classifier: <mrow><msub><mi>o</mi><mi>k</mi></msub><mo>=</mo><mfrac><mrow><msub><mi>CN</mi><mi>k</mi></msub></mrow><mi>e</mi></mfrac><mo>,</mo></mrow> Among them, CN _k is the number of correctly classified test samples on the kth nearest neighbor classifier, and o _k is the recognition rate of the kth nearest neighbor classifier; (8b) Calculate the linear weight coefficients α _k of the K nearest neighbor classifiers respectively according to the recognition rates of the K nearest neighbor classifiers obtained in step (8a): <mrow><msub><mi>&amp;alpha;</mi><mi>k</mi></msub><mo>=</mo><mfrac><msub><mi>o</mi><mi>k</mi></msub><mrow><munderover><mo>&amp;Sigma;</mo><mrow><mi>k</mi><mo>=</mo><mn>1</mn></mrow><mi>K</mi></munderover><msub><mi>o</mi><mi>k</mi></msub></mrow></mfrac><mo>;</mo></mrow> (8c) According to the linear weight coefficient α _k obtained in step (8b), calculate the test sample K channel projection coefficient vectors The weighted distance between the K channel low-rank robust features h _j (k) of the training sample: <mrow><msub><mi>d</mi><mrow><mi>j</mi><mi>i</mi></mrow></msub><mo>=</mo><msub><mi>&amp;alpha;</mi><mn>1</mn></msub><mo>|</mo><mo>|</mo><msub><mover><mi>h</mi><mo>^</mo></mover><mi>i</mi></msub><mrow><mo>(</mo><mn>1</mn><mo>)</mo></mrow><mo>-</mo><msub><mi>h</mi><mi>j</mi></msub><mrow><mo>(</mo><mn>1</mn><mo>)</mo></mrow><mo>|</mo><msub><mo>|</mo><mn>2</mn></msub><mo>+</mo><msub><mi>&amp;alpha;</mi><mn>2</mn></msub><mo>|</mo><mo>|</mo><msub><mover><mi>h</mi><mo>^</mo></mover><mi>i</mi></msub><mrow><mo>(</mo><mn>2</mn><mo>)</mo></mrow><mo>-</mo><msub><mi>h</mi><mi>j</mi></msub><mrow><mo>(</mo><mn>2</mn><mo>)</mo></mrow><mo>|</mo><msub><mo>|</mo><mn>2</mn></msub><mo>+</mo><mo>...</mo><mo>+</mo><msub><mi>&amp;alpha;</mi><mi>k</mi></msub><mo>|</mo><mo>|</mo><msub><mover><mi>h</mi><mo>^</mo></mover><mi>i</mi></msub><mrow><mo>(</mo><mi>k</mi><mo>)</mo></mrow><mo>-</mo><msub><mi>h</mi><mi>j</mi></msub><mrow><mo>(</mo><mi>k</mi><mo>)</mo></mrow><mo>|</mo><msub><mo>|</mo><mn>2</mn></msub><mo>+</mo><mo>...</mo><mo>+</mo><msub><mi>&amp;alpha;</mi><mi>K</mi></msub><mo>|</mo><mo>|</mo><msub><mover><mi>h</mi><mo>^</mo></mover><mi>i</mi></msub><mrow><mo>(</mo><mi>K</mi><mo>)</mo></mrow><mo>-</mo><msub><mi>h</mi><mi>j</mi></msub><mrow><mo>(</mo><mi>K</mi><mo>)</mo></mrow><mo>|</mo><msub><mo>|</mo><mn>2</mn></msub><mo>,</mo></mrow> Get the weighted distance set {d _1i ,d _2i ,...,d _ji ,...,d _ni }, where j=1,2,...,n, i∈{1,2,... ,e}; (8d) According to the weighted distance set {d _1i ,d _2i ,...,d _ji ,...,d _ni } obtained in step (8c), train the ωth corresponding to the minimum value d _ωi in the weighted distance set The category of the sample is the classification result of the test sample, where ω∈{1,2,...,n}.