摘要
多模态人脸识别对于人脸信息挖掘具有重要意义,但目前的多模态人脸快速识别方法的识别率低,误差率高,无法满足快速识别人脸的要求。为了解决该问题,设计基于深度学习特征的多模态人脸快速识别方法。建立人脸特征提取网络,利用几何、模型、统计三个相互并行的卷积神经网络通道,结合深度学习特征分析算法建立卷积神经网络模型,提取人脸特征,由池化操作得到三条特征曲线,并对人脸特征点进行定位。得到人脸特征点后,利用贝塞尔人脸模型建立3D模型,融合卷积神经网络提取到的人脸特征,得到特征向量,利用联合贝叶斯算法计算两个独立的高斯变量,快速识别多模态人脸。实验结果表明,基于深度学习特征的多模态人脸快速识别方法的识别率高达95%,误差率极低,能够在短时间内完成多模态人脸识别工作。
Multimodal face recognition is of great significance for face information mining,but the current multimodal face recognition methods have low recognition rate and high error rate,which can not meet the requirements of rapid face recognition.In order to solve this problem,a new multimodal face recognition method based on deep learning feature is designed.A face feature extraction network is established.Three mutually parallel convolution neural network channels of geometry,model and statistics are used to establish a convolution neural network model combined with the deep learning feature analysis algorithm to extract face features.Three feature curves are obtained by pooling operation,and the face feature points are located.After the facial feature points are obtained,the 3D model is established by using Bessel face model,and the facial features extracted by convolution neural network are fused to obtain the feature vector.The joint Bayesian algorithm is used to calculate two independent Gaussian variables to quickly identify multimodal faces.The experimental results show that the recognition rate of the multimodal face recognition method based on deep learning feature is as high as 95%,and the error rate is very low.It can complete multimodal face recognition in a short time.
作者
雷燕
李杰
董博
孙艳
袁敬
LEI Yan;LI Jie;DONG Bo;SUN Yan;YUAN Jing(CPC State Grid GanSu Electric Power Corporation Party School(State Training Center),Lanzhou 730070,China)
出处
《电子设计工程》
2024年第3期181-184,189,共5页
Electronic Design Engineering
基金
国家电网有限公司科技项目资助(SGGSPX00HLWJS2200097)。
关键词
深度学习特征
多模态人脸
人脸识别
快速识别
联合贝叶斯算法
deep learning feature
multimodalface
face recognition
fastrecognition
recognition research
joint Bayesian algorithm
