摘要
针对人脸面部特征被遮挡后识别率降低的问题,提出一种改进的有无遮挡混合式人脸识别的轻量级MViT模型。改进的MViT模型在MViT Block层加入了三次移位卷积进行局部特征提取,并用GELU激活函数替代RELU;在卷积层加入Flip_merge模块进行人脸面部翻转融合处理;采用Arcface损失函数实现人脸分类识别任务。通过在不同遮挡比例的数据集中进行实验,并在同一数据集上做模块消融实验,该模型的识别率均达到96.5%以上。
Aiming at the problem that the recognition rate of facial features is reduced after occludation,an improved lightweight MViT model for hybrid face recognition with or without occludation is proposed.In the improved MViT model,three⁃shift convolution is added to MViT Block layer for local feature extraction,and GELU activation function is used to replace RELU.Flip_merge module is added to the convolution layer for face and face flip fusion processing.Arcface loss function is used to realize face classification and recognition task.Through experiments on data sets with different occlusion ratios,and module ablation experiments on the same data set,the recognition rate of the model reaches more than 96.5%.
作者
李哲
苗丹
王承林
高娇娇
贾晓雪
LI Zhe;MIAO Dan;WANG Chenglin;GAO Jiaojiao;JIA Xiaoxue(College of Physics and Electronic Engineering,Xingtai University,Xingtai 054000,China;Functional Polymer Materials R&D and Engineering Application Technology Innovation Center of Hebei,Xingtai 054000,China;Key Laboratory of Robot Intelligent Inspection and Sorting Technology Research and Application of Xingtai,Xingtai 054000,China)
出处
《电子设计工程》
2024年第21期156-161,共6页
Electronic Design Engineering
基金
河北省高等学校科学研究项目(ZC2024032)
国家级大学生训练项目(202410104007)。
关键词
人脸识别
特征提取
轻量级
MViT
face recognition
feature extraction
lightweight
MViT
