摘要
传统的卷积神经网络(CNN)在人脸识别中应用极为广泛,然而依然存在收敛速度慢的问题,需要进行批归一化,防止梯度弥散.而自归一化卷积神经网络比普通卷积神经网络收敛速度更快,且无需进行批归一化.因此,提出采用自归一化卷积神经网络来进行人脸识别.首先算法由2个卷积层,1个池化层,2个全连接层和1个Softmax回归层组成的自归一化卷积神经网络对人脸特征进行提取并分类;然后通过对不同批次大小和不同网络层数的实验对比找出最佳的实验条件;最后与传统CNN算法和其他算法对比.提出的方法在ORL数据库中的实验识别率可达到98.3%.实验结果表明,自归一化卷积神经网络比普通的卷积神经网络在人脸识别中具有更高的识别率、更快的收敛速度.
The traditional convolution neural network is widely used in face recognition, but there is still a slow convergence rate.It needs to processbatch of normalization,and prevent gradient diffusion.And then the self- normalizing convolution neural network than the ordinary convolution neural network convergence speed is faster, and no need toproeess batch of normalization.Therefore, it proposes that using the self-normalizing convolution neural network to process face recognition. Firstly, the algorithmstructures a self-normalized convolution neural network consisting of two convolution layers, a pooling layer, two fully connected layers and a Softmax regression layer to process extraction and classification for facial features;then find the best experimental conditions by means of self-normalized convolution neural networks with different batch sizes and different network layers ex- perimental comparison;finally,compared with the traditional CNN algorithm and other algorithms.The recognition rate made by the proposedmethod can reach 98.3% on ORL face database.The experimental results show that the- self-normalizing convolution neural network is more suitable for face recognition than ordinary convolution neural networks, and has higher recognition rate and faster convergence speed.
作者
邱爽
聂仁灿
周冬明
李兴叶
QIU Shuang;NIE Ren-can;ZHOU Dong-ming;LI Xing-ye(School of Information Science and Engineering,Yunnan University, Kunming 650500, Chin)
出处
《云南大学学报（自然科学版）》
CAS
CSCD
北大核心
2018年第4期659-664,共6页
Journal of Yunnan University(Natural Sciences Edition)
基金
国家自然科学基金(61463052
61365001)
关键词
卷积神经网络
自归一化
人脸识别
梯度弥散
特征提取
convolution neural network
self-normalizing
face recognition
gradient dispersion
feature extraction
