Apparatus and method for training classifying model

1. An apparatus for training a classifying model, comprising:

a memory; and

a processor coupled to the memory, wherein the processor is configured to:

input each sample image in a training set to a first machine learning framework, to obtain a first classification probability and a first classification loss of the sample image;

input a second image of an entity, to which the sample image belongs, to a second machine learning framework to obtain a second classification probability and a second classification loss of the second image, wherein the first machine learning framework and the second machine learning framework have identical structures and share identical parameters, and wherein the sample image and the second image corresponding to the sample image belong to a same class;

calculate a similarity loss related to a similarity between the first classification probability and the second classification probability;

calculate, for all sample images in the training set, the sum of the similarity loss, the first classification loss and the second classification loss obtained through calculation with respect to each sample image, as a total loss; and

adjust parameters of the first machine learning framework and the second machine learning framework in a manner of optimizing the total loss, to obtain a trained classifying model.

2. The apparatus for training a classifying model according to claim 1 , wherein the similarity loss is smaller if the similarity between the first classification probability and the second classification probability is higher, and the similarity loss is greater if the similarity between the first classification probability and the second classification probability is lower.

3. The apparatus for training a classifying model according to claim 1 , wherein the sample image is an unobstructed image, and the second image is an image obtained by performing predetermined processing on the sample image.

4. The apparatus for training a classifying model according to claim 3 , wherein the sample image is a face image of a human, and the second image is an image including eyes which is obtained by performing predetermined processing on the sample image.

5. The apparatus for training a classifying model according to claim 4 , wherein the second image is an obstructed image obtained by obstructing forehead hair in the sample image, or the second image is an image obtained by removing glasses in the sample image.

6. The apparatus for training a classifying model according to claim 3 , wherein the second image is an image obtained by performing random fuzzy processing on the sample image.

7. The apparatus for training a classifying model according to claim 1 , wherein the sample image is a front image of a face of a human and the second image is obtained by performing predetermined processing on images of different postures of a human to which the sample image belongs.

8. The apparatus for training a classifying model according to claim 7 , wherein the second image is obtained by performing affine transformation on the images of different postures.

9. The apparatus for training a classifying model according to claim 1 , wherein the similarity loss is calculated using a transfer loss function between the first classification probability and the second classification probability, the transfer loss function comprising one of a cross entropy function, a KL divergence function and a maximum mean difference function.

10. The apparatus for training a classifying model according to claim 1 , wherein before the similarity loss related to the similarity between the first classification probability and the second classification probability is calculated, a softening parameter causing information between different classes to be held is applied to the first classification probability and the second classification probability, respectively, to soften the first classification probability and the second classification probability.

11. An apparatus for performing classification by using the trained classifying model obtained by the apparatus for training the classifying model according to claim 1 , comprising:

a memory; and

a processor coupled to the memory, wherein the processor is configured to classify images to be classified by inputting the images to be classified to the first machine learning framework or the second machine learning framework.

12. A method for training a classifying model, comprising:

inputting each sample image in a training set to a first machine learning framework, to obtain a first classification probability and a first classification loss of the sample image;

inputting a second image of an entity, to which the sample image belongs, to a second machine learning framework to obtain a second classification probability and a second classification loss of the second image, wherein the first machine learning framework and the second machine learning framework have identical structures and share identical parameters, and wherein the sample image and the second image corresponding to the sample image belong to a same class;

calculating a similarity loss related to a similarity between the first classification probability and the second classification probability;

calculating, for all sample images in the training set, the sum of the similarity loss, the first classification loss and the second classification loss obtained through calculation with respect to each sample image, as a total loss; and

adjusting parameters of the first machine learning framework and the second machine learning framework in a manner of optimizing the total loss, to obtain a trained classifying model.

13. The method for training a classifying model according to claim 12 , wherein the similarity loss is smaller if the similarity between the first classification probability and the second classification probability is higher, and the similarity loss is greater if the similarity between the first classification probability and the second classification probability is lower.

14. The method for training a classifying model according to claim 12 , wherein the sample image is an unobstructed image, and the second image is an image obtained by performing predetermined processing on the sample image.

15. The method for training a classifying model according to claim 14 , wherein the sample image is a face image of a human, and the second image is an image including eyes which is obtained by performing predetermined processing on the sample image.

16. The method for training a classifying model according to claim 15 , wherein the second image is an obstructed image obtained by obstructing forehead hair in the sample image, or the second image is an image obtained by removing glasses in the sample image.

17. The method for training a classifying model according to claim 14 , wherein the second image is an image obtained by performing random fuzzy processing on the sample image.

18. The method for training a classifying model according to claim 12 , wherein the sample image is a front image of a face of a human, and the second image is obtained by performing predetermined processing on images of different postures of a human to which the sample image belongs.

19. The method for training a classifying model according to claim 18 , wherein the second image is obtained by performing affine transformation on the images of different postures.

20. The method for training a classifying model according to claim 12 , wherein the similarity loss is calculated using a transfer loss function between the first classification probability and the second classification probability, the transfer loss function comprising one of a cross entropy function, a KL divergence function and a maximum mean difference function.