Method for training feature extraction model, method for classifying image, and related apparatuses

Patent metadata

Field	Value
Publication number	US-12380681-B2
Application number	US-202318183590-A
Country	US
Kind code	B2
Filing date	Mar 14, 2023
Priority date	Sep 29, 2022
Publication date	Aug 5, 2025
Grant date	Aug 5, 2025

Abstract

The present disclosure provides a method for training a feature extraction model, a method for classifying an image and related apparatuses, and relates to the field of artificial intelligence technology such as deep learning and image recognition. The scheme comprises: extracting an image feature of each sample image in a sample image set using a basic feature extraction module of an initial feature extraction model, to obtain an initial feature vector set; performing normalization processing on each initial feature vector in the initial feature vector set using a normalization processing module of the initial feature extraction model, to obtain each normalized feature vector; and guiding training for the initial feature extraction model through a preset high discriminative loss function, to obtain a target feature extraction model as a training result.

First claim

What is claimed is:

1. A method for training a feature extraction model, comprising:
extracting an image feature of each sample image in a sample image set using a basic feature extraction module of an initial feature extraction model, to obtain an initial feature vector set, wherein the sample image set contains sample images of a plurality of image categories, and the each sample image is annotated with an image category to which the sample image belongs;
performing normalization processing on each initial feature vector in the initial feature vector set using a normalization processing module of the initial feature extraction model, to obtain each normalized feature vector; and
guiding training for the initial feature extraction model through a preset high discriminative loss function, to obtain a target feature extraction model as a training result, wherein the high discriminative loss function is abstractly obtained based on a preset upper limit and a preset lower limit of a vector similarity that are preset respectively for sample images of any two image categories according to whether two images have a similarity, a vector similarity between normalized feature vectors of two images not having a similarity is not greater than the preset lower limit, a vector similarity between normalized feature vectors of two images having a similarity is not less than the preset upper limit, and the preset upper limit is greater than the preset lower limit.

2. The method according to claim 1 , wherein the guiding training for the initial feature extraction model through a preset high discriminative loss function comprises:
obtaining guidance information by using the high discriminative loss function to guide a normalization processing process of the normalization processing module for an initial feature vector; and
controlling the normalization processing module to guide an image feature extraction process of the basic feature extraction module in reverse by the guidance information.

3. The method according to claim 2 , further comprising:
using, in response to the basic feature extraction module containing a plurality of feature extraction sub-modules connected in sequence, the high discriminative loss function additionally to guide a feature extraction process of at least one of the feature extraction sub-modules.

4. The method according to claim 1 , wherein a similarity difference between the preset upper limit and the preset lower limit is not less than half of a complete similarity interval.

5. The method according to claim 1 , further comprising:
acquiring a to-be-classified image;
obtaining an outputted actual normalized feature vector by inputting the to-be-classified image into the target feature extraction model;
calculating respectively a vector similarity between the actual normalized feature vector and a standard normalized feature vector of each image category in a feature vector library; and
determining an image category to which the to-be-classified image belongs according to a size of the vector similarity.

6. The method according to claim 5 , wherein the determining an image category to which the to-be-classified image belongs according to a size of the vector similarity comprises:
determining a target standard normalized feature vector having a maximum vector similarity to the actual normalized feature vector; and
determining, in response to the maximum vector similarity being not less than the preset upper limit, an image category to which the target standard normalized feature vector belongs as the image category to which the to-be-classified image belongs.

7. The method according to claim 6 , further comprising:
returning, in response to the maximum vector similarity being less than the preset upper limit, an image classification abnormality notification of failing to determine the image category to which the to-be-classified image belongs.

8. The method according to claim 7 , further comprising:
using a to-be-classified image corresponding to the returned image classification abnormality notification as a newly added image category; and
using the actual normalized feature vector as a standard normalized feature vector under the newly added image category to supplement the feature vector library.

9. An electronic device, comprising:
at least one processor; and
a storage device, in communication with the at least one processor,
wherein the storage device stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor, to enable the at least one processor to perform operations for training a feature extraction model, the operations comprising:
extracting an image feature of each sample image in a sample image set using a basic feature extraction module of an initial feature extraction model, to obtain an initial feature vector set, wherein the sample image set contains sample images of a plurality of image categories, and the each sample image is annotated with an image category to which the sample image belongs;
performing normalization processing on each initial feature vector in the initial feature vector set using a normalization processing module of the initial feature extraction model, to obtain each normalized feature vector; and
guiding training for the initial feature extraction model through a preset high discriminative loss function, to obtain a target feature extraction model as a training result, wherein the high discriminative loss function is abstractly obtained based on a preset upper limit and a preset lower limit of a vector similarity that are preset respectively for sample images of any two image categories according to whether two images have a similarity, a vector similarity between normalized feature vectors of two images not having a similarity is not greater than the preset lower limit, a vector similarity between normalized feature vectors of two images having a similarity is not less than the preset upper limit, and the preset upper limit is greater than the preset lower limit.

10. The electronic device according to claim 9 , wherein the guiding training for the initial feature extraction model through a preset high discriminative loss function comprises:
obtaining guidance information by using the high discriminative loss function to guide a normalization processing process of the normalization processing module for an initial feature vector; and
controlling the normalization processing module to guide an image feature extraction process of the basic feature extraction module in reverse by the guidance information.

11. The electronic device according to claim 10 , the operations further comprising:
using, in response to the basic feature extraction module containing a plurality of feature extraction sub-modules connected in sequence, the high discriminative loss function additionally to guide a feature extraction process of at least one of the feature extraction sub-modules.

12. The electronic device according to claim 9 , wherein a similarity difference between the preset upper limit and the preset lower limit is not less than half of a complete similarity interval.

13. The electronic device according to claim 9 , the operations further comprising:
acquiring a to-be-classified image;
obtaining an outputted actual normalized feature vector by inputting the to-be-classified image into the target feature extraction model;
calculating respectively a vector similarity between the actual normalized feature vector and a standard normalized feature vector of each image category in a feature vector library; and
determining an image c

Assignees

• Beijing Baidu Netcom Sci & Tech Co Ltd

Inventors

• Dong Shuilong
• He Sensen
• Wei Shengyu
• Cui Cheng
• Du Yuning
• Gao Tingquan
• Zeng Shao
• Zhou Ying
• Lyu Xueying
• Liu Yi
• Zhao Qiao
• Liu Qiwen
• Bi Ran
• Hu Xiaoguang
• Yu Dianhai
• Ma Yanjun

Classifications

• G06V10/761

  Proximity, similarity or dissimilarity measures · CPC title

• G06V10/776

  Validation; Performance evaluation · CPC title

• G06V10/764

  using classification, e.g. of video objects · CPC title

• G06V10/7784

  based on feedback from supervisors · CPC title

• G06V10/40

  Extraction of image or video features · CPC title