Training method and device for image identifying model, and image identifying method

Patent metadata

Field	Value
Publication number	US-12106545-B2
Application number	US-202117534681-A
Country	US
Kind code	B2
Filing date	Nov 24, 2021
Priority date	Jul 29, 2021
Publication date	Oct 1, 2024
Grant date	Oct 1, 2024

Abstract

The present disclosure provides a training method and device for an image identifying model, and an image identifying method. The training method comprises: obtaining image samples of a plurality of categories; inputting image samples of each category into a feature extraction layer of the image identifying model to extract a feature vector of each image sample; calculating a statistical characteristic information of an actual distribution function corresponding to each category according to the feature vector of each image sample of the each category; establishing an augmented distribution function corresponding to the each category according to the statistical characteristic information; obtaining augmented sample features of the each category based on the augmented distribution function; and inputting feature vectors of the image samples and the augmented sample features into a classification layer of the image identifying model for supervised learning.

First claim

What is claimed is:

1. A training method for an image identifying model, comprising:
obtaining image samples of a plurality of categories;
inputting image samples of each category of the plurality of categories into a feature extraction layer of the image identifying model to extract a feature vector of each image sample;
calculating a statistical characteristic information of an actual distribution function corresponding to each category according to the feature vector of each image sample of the each category;
establishing an augmented distribution function corresponding to the each category according to the statistical characteristic information of the actual distribution function corresponding to the each category;
obtaining augmented sample features of the each category based on the augmented distribution function corresponding to the each category; and
inputting feature vectors of the image samples and the augmented sample features into a classification layer of the image identifying model for supervised learning,
wherein the statistical characteristic information comprises a first statistical characteristic information and a second statistical characteristic information; and
the establishing of the augmented distribution function corresponding to the each category comprises:
calculating an average value of the second statistical characteristic information of actual distribution functions corresponding to the plurality of categories; and
establishing the augmented distribution function corresponding to the each category in a case where the first statistical characteristic information of the each category and the average value of the second statistical characteristic information are used as statistical characteristic parameters.

2. The training method according to claim 1 , wherein the first statistical characteristic information is a mean value of the actual distribution function, and the second statistical characteristic information is a variance or a standard deviation of the actual distribution function.

3. The training method according to claim 1 , wherein the obtaining of the image samples of the plurality of categories comprises:
collecting image samples of the plurality of categories under a same attribute to obtain at least a portion of image samples of each category.

4. The training method according to claim 3 , wherein numbers of image samples of all categories collected in the plurality of categories are equal.

5. The training method according to claim 3 , wherein the collecting of the image samples of the plurality of categories under the same attribute comprises:
collecting repeatedly at least part of image samples of a first category in the plurality of categories in a case where an actual number of image samples in the first category is less than a planned number of image samples collected from the first category.

6. The training method according to claim 3 , wherein the collecting of the image samples of the plurality of categories under the same attribute comprises:
collecting all image samples of a second category in the plurality of categories in a case where an actual number of image samples in the second category is equal to a planned number of image samples collected from the second category.

7. The training method according to claim 3 , wherein the collecting of the image samples of the plurality of categories under the same attribute comprises:
collecting a part of image samples of a third category in the plurality of categories in a case where an actual number of image samples in the third category is greater than a planned number of image samples collected from the third category.

8. The training method according to claim 1 , wherein the obtaining of the augmented sample features of the each category based on the augmented distribution function corresponding to the each category comprises:
sampling the augmented distribution function corresponding to the each category to obtain the augmented sample features corresponding to the each category.

9. The training method according to claim 8 , wherein a number of the image samples of the each category is equal to a number of the augmented sample features of the each category.

10. The training method according to claim 1 , wherein the actual distribution function and the augmented distribution function are Gaussian distribution functions.

11. The training method according to claim 3 , wherein:
the attribute represents a certain part of a face;
the category is a shape category of the certain part of the face; and
the image sample comprises an image sample of the certain part of the face.

12. The training method according to claim 1 , wherein a function type of the augmented distribution function is the same as a function type of the actual distribution function.

13. The training method according to claim 1 , wherein the image identifying model is a convolutional neural network.

14. An image identifying method, comprising:
inputting an image to be identified into an image identifying model, wherein the image identifying model is trained by the training method according to claim 1 ; and
identifying the image to be identified and outputting an image identifying result by the image identifying model.

15. A training device for an image identifying model, comprising:
a first obtaining unit configured to obtain image samples of a plurality of categories;
a feature extraction unit configured to input image samples of each category of the plurality of categories into a feature extraction layer of the image identifying model to extract a feature vector of each image sample;
a calculating unit configured to calculate a statistical characteristic information of an actual distribution function corresponding to each category according to the feature vector of each image sample of the each category;
an augmented distribution function establishing unit configured to establish an augmented distribution function corresponding to the each category according to the statistical characteristic information of the actual distribution function corresponding to the each category;
a second obtaining unit configured to obtain augmented sample features of the each category based on the augmented distribution function corresponding to the each category; and
a supervised learning unit configured to input feature vectors of the image samples and the augmented sample features into a classification layer of the image identifying model for supervised learning,
wherein the statistical characteristic information comprises a first statistical characteristic information and a second statistical characteristic information; and
the augmented distribution function establishing unit is configured to calculate an average value of the second statistical characteristic information of actual distribution functions corresponding to the plurality of categories, and establish the augmented distribution function corresponding to the each category in a case where the first statistical characteristic information of the each category and the average value of the second statistical characteristic information are used as statistical characteristic parameters.

16. An image identifying device, comprising:
an input unit configured to input an image to be identified into an image identifying model, wherein the image identifying model is trained by the training method according to claim 1 ; and
an identification unit configured to identify the image to be identified and output an image identifying result by the image identifying model.

Assignees

• Lemon Inc

Inventors

• Sun Jingna
• Chen Peibin
• Zeng Weihong
• Wang Xu
• Liu Jing
• Lai Chunpong
• Sang Shen

Classifications

• G06V10/72 Primary

  Data preparation, e.g. statistical preprocessing of image or video features · CPC title

• G06V10/82

  using neural networks · CPC title

• G06N3/08

  Learning methods · CPC title

• G06N3/09

  Supervised learning · CPC title

• G06N3/0464

  Convolutional networks [CNN, ConvNet] · CPC title