Lightweight face recognition method based on DenseNet dual-module network structure

By adopting the DenseNet dual-module network structure and angle weighting mechanism in the face detection method, the deployment difficulties and accuracy instability caused by the large amount of model parameters in the prior art are solved, and an efficient and lightweight face recognition system is realized.

CN119693985APending Publication Date: 2025-03-25ZHEJIANG UNIV OF TECH
2 Cites 1 Cited by

Patent Information

Application Number
CN202411751574.2
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-12-02
Publication Date
2025-03-25

AI Technical Summary

Technical Problem

The existing face detection method based on deep learning is difficult to deploy and unstable accuracy due to the large amount of model parameters, especially in resource-constrained environments.

Method used

A lightweight face recognition method based on DenseNet dual-module network structure is adopted. By adjusting the main network structure, DenseNet Block is introduced to replace MobileVit Block, and an angle weighting mechanism is introduced based on the traditional L2 loss function to enhance the penalty coefficient of side face images.

Benefits of technology

While ensuring high detection accuracy, it significantly reduces the amount of model parameters, improves computing efficiency, reduces memory usage and computing consumption, and is suitable for deployment in resource-constrained environments.

✦ Generated by Eureka AI based on patent content.
Patent Text Reader

Abstract

A lightweight face recognition method based on a DenseNet dual-module network structure uses a lightweight network DenseNet as a backbone network and attitude angle feedback information as an auxiliary network, and introduces an angle punishment mechanism; the auxiliary network and the angle penalty mechanism combine angle loss with key point detection loss in the training process to help the main network to better learn the relationship between face key points. According to the method provided by the invention, a contrast experiment is carried out on a disclosed WFLW data set, so that relatively high detection precision is ensured, and the quantity of model parameters is greatly reduced.
Need to check novelty before this filing date? Find Prior Art