Image recognition method and apparatus

• US 10,713,532 B2
• Filed: 03/19/2018
• Issued: 07/14/2020
• Est. Priority Date: 11/26/2015
• Status: Active Grant

• Current Assignee: Tencent Technology Company Limited (Tencent Holdings Limited)
• Original Assignee: Tencent Technology Company Limited (Tencent Holdings Limited)
• Inventors: Ding, Shouhong, Li, Jilin, Wang, Chengjie, Huang, Feiyue, Wu, Yongjian, Tan, Guofu
• Primary Examiner(s): Saini, Amandeep
• Application Number:
  US15/925,028
• Publication Number:
  US 20180204094A1
• Time in Patent Office: 848 Days
• Field of Search: 382118
• CPC Class Codes: G06F 18/2135 based on approximation criteria, e.g. principal component analysis
• CPC Class Codes: G06F 18/214 Generating training patterns; Bootstrap methods, e.g. bagging or boosting
• CPC Class Codes: G06F 18/24155 Bayesian classification
• CPC Class Codes: G06V 10/467 Encoded features or binary features, e.g. local binary patterns [LBP]
• CPC Class Codes: G06V 10/764 using classification, e.g. of video objects
• CPC Class Codes: G06V 10/7715 Feature extraction, e.g. by transforming the feature space, e.g. multi-dimensional scaling [MDS]; Mappings, e.g. subspace methods
• CPC Class Codes: G06V 40/165 using facial parts and geometric relationships
• CPC Class Codes: G06V 40/168 Feature extraction; Face representation
• CPC Class Codes: G06V 40/171 Local features and components; Facial parts eye characteristics G06V40/18; Occluding parts, e.g. glasses; Geometrical relationships
• CPC Class Codes: G06V 40/172 Classification, e.g. identification

Abstract

The present disclosure discloses an image recognition method and apparatus, and belongs to the field of computer technologies. The method includes: extracting a local binary pattern (LBP) feature vector of a target image; calculating a high-dimensional feature vector of the target image according to the LBP feature vector; obtaining a training matrix, the training matrix being a matrix obtained by training images in an image library by using a joint Bayesian algorithm; and recognizing the target image according to the high-dimensional feature vector of the target image and the training matrix. The image recognition method and apparatus according to the present disclosure may combine LBP algorithm with a joint Bayesian algorithm to perform recognition, thereby improving the accuracy of image recognition.

Claims

1. An image recognition method, comprising:

extracting a local binary pattern (LBP) feature vector of a target image, the target image being a face image;

calculating a high-dimensional feature vector of the target image according to the LBP feature vector;

obtaining a training matrix, the training matrix being a matrix obtained by training images in an image library by using a joint Bayesian algorithm, the image library comprising m*n face images, m being a number of persons corresponding to images in the image library, and n being a number of face images of each person; and

performing image recognition on the target image according to the high-dimensional feature vector of the target image and the training matrix,

wherein obtaining the training matrix comprises;

obtaining a high-dimensional feature vector of each person corresponding to the face images in the image library;

initializing a first covariance matrix of n face images of each person in the image library and a second covariance matrix of images of different persons in the image library, the first covariance matrix being an m-dimensional square matrix and the second covariance matrix being an m*n square matrix;

calculating a mean value of Gaussian distribution of each person in the image library;

calculating a joint distribution covariance matrix of two persons in the image library according to the first covariance matrix, the second covariance matrix, and the high-dimensional feature vectors of the two persons;

updating the first covariance matrix according to the mean value of Gaussian distribution, and updating the second covariance matrix according to the joint distribution covariance matrix; and

obtaining the training matrix based on the updated first covariance matrix and the updated second covariance matrix.

2. The method according to claim 1, wherein performing image recognition on the target image according to the high-dimensional feature vector of the target image and the training matrix comprises:

obtaining a high-dimensional feature vector of a reference image;

calculating a similarity between the target image and the reference image according to the high-dimensional feature vector of the target image, the high-dimensional feature vector of the reference image, and the training matrix;

r(x₁,x₂)=x₁^TAx₁+x₂^TAx₂−

2x₁^TGx₂;

wherein x₁ is the high-dimensional feature vector of the target image, x₂ is the high-dimensional feature vector of the reference image, x₁^T a transposed vector of x₁, x₂^T is a transposed vector of x₂, and A and G are training matrixes.

3. The method according to claim 2, wherein the image library comprises the target image and the reference image.

4. The method according to claim 1, wherein the extracting an LBP feature vector of a target image comprises:

separately scaling the target image according to k preset scaling factors to obtain k scaled images, k being a positive integer;

determining, for each of the target image and the k scaled images, a key point in the image, and extracting an LBP feature vector at the key point; and

determining the LBP feature vector of the target image according to the extracted LBP feature vector of the key point.

5. The method according to claim 1, wherein the calculating a high-dimensional feature vector of the target image according to the LBP feature vector comprises:

performing principal component analysis (PCA) dimension reduction on x_r^Tx_r, and reserving features of first N dimensions, to obtain a reduced-dimension matrix P, N being a positive integer, x_r being the LBP feature vector of the target image, and x_r^T being a transposed vector of x_r;

for x_r, calculating X_p=Px_r;

performing latent Dirichlet allocation (LDA) dimension reduction on x_p to obtain a reduced-dimension matrix L; and

calculating the high-dimensional feature vector of the target image x=Lx_p.

6. The method according to claim 1, wherein obtaining the training matrix, further comprises:

initializing the first covariance matrix S_μ

of n face images of each person in the image library and the second covariance matrix S_ε

of images of different persons in the image library;

calculating F and G;

F=S_μ

<sup>−

1</sup>, G=−

(mS_μ

+S_ε

)<sup>−

1</sup>S_μ

S_ε

<sup>−

1</sup>;

calculating the mean value μ

_i of Gaussian distribution of an i^th person in the image library;

μ

_i=Σ

_i=1^m S_μ

(F+mG)x_i, and the joint distribution covariance matrix ε

_ij of the i^th person and a j^th person according to F and G;

ε

_ij=x_j+Σ

_i=1^m S_ε

Gx_i, x_i being a high-dimensional feature vector of the i^th person, and x_j being a high-dimensional feature vector of the j^th person;

updating S_μ

and S_ε

according to the following formula;

7. The method according to claim 6, further comprising:

determining whether the updated S_μ

and S_ε

are convergent; and

if the updated S_μ

and S_ε

are not convergent, performing the step of calculating F and G again;

or

if the updated S_μ

and S_ε

are convergent, performing a step of calculating the training matrix according to F and G corresponding to the convergent S_μ

and S_ε

, and the convergent S_μ

and S_ε

.

8. The method according to claim 1, wherein the obtaining a high-dimensional feature vector of each person corresponding to the images in the image library comprises:

extracting, for each face image in the image library, an LBP feature vector of the face image;

calculating a mean value of all features in the LBP feature vector of the face image;

subtracting the mean value from each feature in the LBP feature vector of the face image to obtain a feature vector x_r={x_ri, 0<

i<

m*n}, wherein x_ri represents an LBP feature vector of the face image after mean value processing;

performing PCA dimension reduction on x_r^Tx_r, and reserving features of the first N dimensions, to obtain a reduced-dimension matrix P, N being a positive integer, and x_r^T being a transposed vector of x_r;

for x_r, calculating x_p=Px_r;

performing, for face images of each person, LDA dimension reduction on x_p to obtain a reduced-dimension matrix L; and

calculating the high-dimensional feature vector of each person;

x=Lx_p.

9. An image recognition apparatus, comprising:

at least one processor; and

a memory, the memory storing program instructions, and the image recognition apparatus being configured to perform the following operations when the program instructions are performed by the processor;

extracting a local binary pattern (LBP) feature vector of a target image, the target image being a face image;

calculating a high-dimensional feature vector of the target image according to the LBP feature vector;

obtaining a training matrix, the training matrix being a matrix obtained by training images in an image library by using a joint Bayesian algorithm, the image library comprising m*n face images, m being a number of persons corresponding to images in the image library, and n being a number of face images of each person; and

performing image recognition on the target image according to the high-dimensional feature vector of the target image and the training matrix,

wherein obtaining the training matrix comprises;

obtaining a high-dimensional feature vector of each person corresponding to the face images in the image library;

initializing a first covariance matrix of n face images of each person in the image library and a second covariance matrix of images of different persons in the image library, the first covariance matrix being an m-dimensional square matrix and the second covariance matrix being an m*n square matrix;

calculating a mean value of Gaussian distribution of each person in the image library;

calculating a joint distribution covariance matrix of two persons in the image library according to the first covariance matrix, the second covariance matrix, and the high-dimensional feature vectors of the two persons;

updating the first covariance matrix according to the mean value of Gaussian distribution, and updating the second covariance matrix according to the joint distribution covariance matrix; and

obtaining the training matrix based on the updated first covariance matrix and the updated second covariance matrix.

10. The apparatus according to claim 9, wherein the performing image recognition on the target image according to the high-dimensional feature vector of the target image and the training matrix comprises:

obtaining a high-dimensional feature vector of a reference image;

calculating a similarity between the target image and the reference image according to the high-dimensional feature vector of the target image, the high-dimensional feature vector of the reference image, and the training matrix;

r(x₁,x₂)=x₁^TAx₁+x₂^TAx₂−

2x₁^TGx₂;

wherein x₁ is the high-dimensional feature vector of the target image, x₂ is the high-dimensional feature vector of the reference image, x₁^T a transposed vector of x₁, x₂^T is a transposed vector of x₂, and A and G are training matrixes.

11. The apparatus according to claim 10, wherein the image library comprises the target image and the reference image.

12. The apparatus according to claim 9, wherein the extracting an LBP feature vector of a target image comprises:

separately scaling the target image according to k preset scaling factors to obtain k scaled images, k being a positive integer;

determining, for each of the target image and the k scaled images, a key point in the image, and extracting an LBP feature vector at the key point; and

determining the LBP feature vector of the target image according to the LBP feature vector of the extracted key point.

13. The apparatus according to claim 9, wherein the calculating a high-dimensional feature vector of the target image according to the LBP feature vector comprises:

performing principal component analysis (PCA) dimension reduction on x_r^Tx_r, and reserving features of first N dimensions, to obtain a reduced-dimension matrix P, N being a positive integer, x_r being the LBP feature vector of the target image, and x_r^T being a transposed vector of x_r;

for x_r, calculating X_p=Px_r;

performing latent Dirichlet allocation (LDA) dimension reduction on x_p to obtain a reduced-dimension matrix L; and

calculating the high-dimensional feature vector of the target image x=Lx_p.

14. The apparatus according to claim 9, wherein obtaining the training matrix further comprises:

initializing the first covariance matrix S_μ

of n face images of each person in the image library and the second covariance matrix S_ε

of images of different persons in the image library;

calculating F and G;

F=S_μ

<sup>−

1</sup>, G=−

(mS_μ

+S_ε

)<sup>−

1</sup>S_μ

S_ε

<sup>−

1</sup>;

calculating the mean value μ

_i of Gaussian distribution of an i^th person in the image library;

μ

_i=Σ

_i=1^m S_μ

(F+mG)x_i, and the joint distribution covariance matrix ε

_ij of the i^th person and a j^th person according to F and G;

ε

_ij=x_j+Σ

_i=1^mS_ε

Gx_i, x_i being a high-dimensional feature vector of the i^th person, and x_j being a high-dimensional feature vector of the j^th person;

updating S_μ

and S_ε

according to the following formula;

15. The apparatus according to claim 14, wherein the operations further comprise:

determining whether the updated S_μ

and S_ε

are convergent; and

if the updated S_μ

and the updated S_ε

are not convergent, performing the step of calculating F and G again;

or

if the updated S_μ

and the updated S_ε

are convergent, performing a step of calculating a training matrix according to F and G corresponding to the convergent S_μ

and S_ε

, and the convergent S_μ

and S_ε

.

16. The apparatus according to claim 9, wherein the obtaining a high-dimensional feature vector of each person corresponding to images in the image library comprises:

extracting, for each face image in the image library, an LBP feature vector of the face image;

calculating a mean value of all features in the LBP feature vector of the face image;

subtracting the mean value from each feature in the LBP feature vector of the face image to obtain a feature vector x_r={x_ri0<

i<

m*n}, wherein x_ri represents an LBP feature vector of the face image after mean value processing;

performing PCA dimension reduction on x_r^Tx_r, and reserving features of the first N dimensions, to obtain a reduced-dimension matrix P, N being a positive integer, and x_r^T being a transposed vector of x_r;

for x_r, calculating x_p=Px_r;

performing, for face images of each person, LDA dimension reduction on x_p to obtain a reduced-dimension matrix L; and

calculating the high-dimensional feature vector of each person;

x=Lx_p.

17. A non-transitory computer-readable storage medium, the storage medium storing program instructions, and when being executed by a processor of a computing device, the program instructions causing the processor to perform:

extracting a local binary pattern (LBP) feature vector of a target image, the target image being a face image;

calculating a high-dimensional feature vector of the target image according to the LBP feature vector;

obtaining a training matrix, the training matrix being a matrix obtained by training images in an image library by using a joint Bayesian algorithm, the image library comprising m*n face images, m being a number of persons corresponding to images in the image library, and n being a number of face images of each person; and

performing image recognition on the target image according to the high-dimensional feature vector of the target image and the training matrix,

wherein obtaining the training matrix comprises;

obtaining a high-dimensional feature vector of each person corresponding to the face images in the image library;

initializing a first covariance matrix of n face images of each person in the image library and a second covariance matrix of images of different persons in the image library, the first covariance matrix being an m-dimensional square matrix and the second covariance matrix being an m*n square matrix;

calculating a mean value of Gaussian distribution of each person in the image library;

calculating a joint distribution covariance matrix of two persons in the image library according to the first covariance matrix, the second covariance matrix, and the high-dimensional feature vectors of the two persons;

updating the first covariance matrix according to the mean value of Gaussian distribution, and updating the second covariance matrix according to the joint distribution covariance matrix; and

obtaining the training matrix based on the updated first covariance matrix and the updated second covariance matrix.