Looking for breakthrough ideas for innovation challenges? Try Patsnap Eureka!

Image classification method, system, medium and device during class distribution mismatching

A classification method and image technology, applied in the field of image processing, can solve the problems that classifiers have no positive effect, affect the accuracy of classification problems, affect the effect of model training, etc., achieve maximum utilization of labeling resources, wide application, and improve training effects Effect

Pending Publication Date: 2021-09-24
RENMIN UNIVERSITY OF CHINA
View PDF5 Cites 0 Cited by
  • Summary
  • Abstract
  • Description
  • Claims
  • Application Information

AI Technical Summary

Problems solved by technology

Mismatched image data has little positive effect on classifier training and can greatly affect the accuracy of classification problems
And the unmatched data is likely to be recognized as meaningful data by the existing active learning framework due to its large difference from the labeled data, thereby crowding out labeling resources and greatly affecting the training effect of the model

Method used

the structure of the environmentally friendly knitted fabric provided by the present invention; figure 2 Flow chart of the yarn wrapping machine for environmentally friendly knitted fabrics and storage devices; image 3 Is the parameter map of the yarn covering machine
View more

Image

Smart Image Click on the blue labels to locate them in the text.
Viewing Examples
Smart Image
  • Image classification method, system, medium and device during class distribution mismatching
  • Image classification method, system, medium and device during class distribution mismatching

Examples

Experimental program
Comparison scheme
Effect test

Embodiment 1

[0021] This embodiment discloses an image classification method when the class distribution does not match, such as including:

[0022] 1. Use the labeled images to train the initial image classification model and verify the image classification model.

[0023] The labeled images here and the unlabeled images below are all preprocessed images. Among them, preprocessing refers to the feature extraction of stored image data. Feature extraction methods include but not limited to directional gradient histogram, scale-invariant feature transformation, neural network feature extraction, etc., and then convert the feature-extracted image into a suitable The form for model training. Usually, the form suitable for model training refers to high-dimensional vectors that can reflect image features.

[0024] 2. If figure 2 As shown, the unlabeled images are selected from the verification results, and their query scores are obtained according to their semantic scores and information sco...

Embodiment 2

[0033] Based on the same inventive concept, this embodiment uses a specific example to describe the solution of Embodiment 1 in detail.

[0034] According to the subject of each image, the images are divided into different categories, for example, the images are divided into three categories: "cat", "dog" and "car". Assuming that during the training of the image classification model, the collected labeled images only contain two categories of "cat" and "dog", but the collected unlabeled images contain data of all three categories, the actual An image of class "car" is the data for which the class distribution does not match. The cycle training of the model is carried out according to the method in Embodiment 1. In each round of cycle, after the image classification model training is performed according to the current marked image, the query score of each item of data in the unlabeled image is calculated. Among them, samples whose real category is "car" will get a lower semant...

Embodiment 3

[0036] Based on the same inventive concept, this embodiment discloses an image classification system when the class distribution does not match, including:

[0037] A pre-training module, which uses marked images to train the initial image classification model and verify the image classification model;

[0038] The query scoring module is used to select unlabeled images from the verification results, obtain their query scores according to their semantic scores and information scores, sort them according to the query scores, and select images with high query scores according to the sorting results for labeling;

[0039] The secondary training module is used to add the marked image to the marked image, and train the initial image classification model until the training result meets the preset requirements;

[0040] An image classification module, which uses an image classification model meeting preset requirements to classify images.

the structure of the environmentally friendly knitted fabric provided by the present invention; figure 2 Flow chart of the yarn wrapping machine for environmentally friendly knitted fabrics and storage devices; image 3 Is the parameter map of the yarn covering machine
Login to View More

PUM

No PUM Login to View More

Abstract

The invention belongs to the technical field of image processing, and relates to an image classification method, system, medium and device during class distribution mismatching, and the method comprises the steps: employing a labeled image to train an initial image classification model, and verifying the image classification model; selecting unlabeled images in the verification result, obtaining query scores of the unlabeled images according to semantic scores and information amount scores of the unlabeled images, sorting the unlabeled images according to the query scores, and selecting the images with the high query scores according to the sorting result to be labeled; adding the marked image into the marked image, and training the initial image classification model until a training result meets a preset requirement; and classifying the images by adopting the image classification model meeting the preset requirement. When distribution of image data has a very high mismatching degree, mismatched unlabeled samples can be effectively avoided in a data selection process, and a relatively high model training capability is kept.

Description

technical field [0001] The invention relates to an image classification method, system, medium and equipment when the class distribution does not match, belongs to the technical field of image processing, and is especially suitable for image classification in active learning. Background technique [0002] Machine learning has become an important research method and analysis method in various fields. As one of the important branches of machine learning, deep learning has made many breakthroughs in the field of supervised learning with sufficient labeled data, and has been widely used in the field of image classification. However, with the gradual increase in the complexity of the data and the increase in the cost of manual labeling, it is often unrealistic to obtain sufficient labeled image data. In order to complete the learning of the classification model under this condition, some weakly supervised learning methods have been proposed and verified to have good performance,...

Claims

the structure of the environmentally friendly knitted fabric provided by the present invention; figure 2 Flow chart of the yarn wrapping machine for environmentally friendly knitted fabrics and storage devices; image 3 Is the parameter map of the yarn covering machine
Login to View More

Application Information

Patent Timeline
no application Login to View More
Patent Type & Authority Applications(China)
IPC IPC(8): G06K9/62G06K9/46
CPCG06F18/241G06F18/214
Inventor 赵素云杜盼甘泽宇陈红李翠平
Owner RENMIN UNIVERSITY OF CHINA
Who we serve
  • R&D Engineer
  • R&D Manager
  • IP Professional
Why Patsnap Eureka
  • Industry Leading Data Capabilities
  • Powerful AI technology
  • Patent DNA Extraction
Social media
Patsnap Eureka Blog
Learn More
PatSnap group products

Browse by: Latest US Patents, China's latest patents, Technical Efficacy Thesaurus, Application Domain, Technology Topic, Popular Technical Reports.

© 2024 PatSnap. All rights reserved.Legal|Privacy policy|Modern Slavery Act Transparency Statement|Sitemap|About US| Contact US: help@patsnap.com