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

A training method combining face recognition data equalization and enhancement

A face recognition and training method technology, applied in the field of image classification based on deep learning, can solve the problems of data difference and real scene, data imbalance, etc., to achieve low loss convergence value, high balance, and fast training convergence. Effect

Active Publication Date: 2020-06-16
CHENGDU KOALA URAN TECH CO LTD
View PDF16 Cites 0 Cited by
  • Summary
  • Abstract
  • Description
  • Claims
  • Application Information

AI Technical Summary

Problems solved by technology

[0007] Aiming at the problems of unbalanced data and differences between data and real scenes in the prior art, the present invention proposes a training scheme combining face recognition data equalization and enhancement. Through online and offline data enhancement and sampling, the human Data balance and enhancement for face recognition

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
  • A training method combining face recognition data equalization and enhancement
  • A training method combining face recognition data equalization and enhancement
  • A training method combining face recognition data equalization and enhancement

Examples

Experimental program
Comparison scheme
Effect test

Embodiment 1

[0043] A training method combining face recognition data equalization and enhancement, such as figure 1 shown, including the following steps:

[0044] S1. Generate a training pre-model using the photos contained in all the cluster label ids of the dataset;

[0045] S2. Screen out the enhanced sample picture bases in all clustering label ids of the data set; the enhanced sample picture bases are pictures whose quality meets certain conditions in each clustering label id;

[0046] S3. Perform offline enhancement on the enhanced sample image base, and filter out the similarity between each cluster label id in the enhanced sample image base after offline enhancement and the cluster center of the cluster label id in the similarity interval (a, b) between n sample pictures;

[0047] S4. Design a data equalization method to perform data equalization on the sample image;

[0048] S5. Perform online enhancement on the sample image after data equalization.

[0049] Working principle...

Embodiment 2

[0051] In order to better realize the present invention, further, in combination with figure 2 As shown, the S2 specifically includes the following steps: S2.1. Use the pre-model generated by S1 to determine the class center of each class of cluster label id of the data set;

[0052] S2.2. According to the cluster center determined by the cluster label id of each category, select the pictures whose similarity with the cluster center is higher than the parameter t in the corresponding cluster label id.

[0053]Working principle: The class center of each cluster label id is determined through the pre-training model. For example, the glint dataset has 180,000 cluster label ids, that is, a matrix of 180000×512 is generated and saved. The calculation formula of each class center is as follows: emb_i=model(img_i)

[0054] mid_emb=(emb_0+emb_1+....emb_n) / nmid_emb.shape:1×512;

[0055] After the class center of each cluster label id is determined through the pre-training model, we ...

Embodiment 3

[0058] In order to better realize the present invention, further, the parameter t can be adjusted according to the size of the data set and actual requirements.

[0059] Working principle: Through the similarity parameter t, the pictures similar to the class center in the data set can be screened out, because the offline enhancement belongs to the category of large-scale enhancement, only the data near the class center is selected for selection, and the similarity parameter t is set for screening , can remove some outlayers and some data at the edge of the class; after many experiments in the present invention, the value of the parameter t is selected to be 0.65, but if the selected data sets are different, and the requirements for the accuracy of the experiment are different, The specific value of the similarity parameter t can be adjusted up and down independently.

[0060] Other parts of this embodiment are the same as those of Embodiment 1-2 above, so they will not be repe...

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

A training method that combines face recognition data balance and enhancement. Firstly, a pre-training model is trained through the data set to achieve the recurring effect; For high-quality pictures in the class label id, the accuracy of face recognition can be increased by screening out the enhanced sample base for recognition; the enhanced sample base is enhanced offline, and the offline enhancement is carried out by selecting a rich and large-scale enhancement degree. It can make the enhanced sample base more diverse, and then filter out the enhanced effect through the similarity interval (a, b), but it will not enhance the picture into an unrecognizable sample; and then design a data balance method, the data Perform equalization so that the number of samples selected by each cluster label id is similar; finally perform an offline enhancement on the balanced data. The present invention realizes the data equalization and enhancement of face recognition through the above steps.

Description

technical field [0001] The invention belongs to the field of image classification based on deep learning, and in particular relates to a training method combining equalization and enhancement of face recognition data. Background technique [0002] In recent years, with the rapid development of computer technology, automatic face recognition technology has been extensively researched and developed, and face recognition has become one of the most popular research topics in pattern recognition and image processing in the past 30 years. Face recognition is a biometric identification technology based on human facial feature information: use a camera or camera to collect images or video streams containing human faces, and automatically detect and track human faces in the images, and then detect A series of related technologies for face recognition. Face recognition is also commonly called portrait recognition and facial recognition. Face recognition technology can compare the ne...

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 Patents(China)
IPC IPC(8): G06K9/62G06T3/00G06T5/00G06K9/00
CPCG06T3/0006G06T5/002G06T5/001G06T2207/20024G06T2207/20032G06T2207/30201G06V40/172G06F18/23213G06F18/214
Inventor 张芮铭沈复民孔繁昊奚兴张艳明
Owner CHENGDU KOALA URAN TECH CO LTD
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