摘要
针对传统神经网络在人脸图像的训练过程中没有将高低卷积层信息进行融合,为充分利用图像各层特征信息,提出一种基于三层特征融合的全连接卷积神经网络模型,算法将原有网络最后三层特征结合,并将提取的特征信息与最后一层全连接层结合,从而增加了浅层特征的表达,加强了深层特征的提取效果,促使改进后的卷积神经网络提取的信息更加完备;同时将损失函数和中心函数加权联合,以提高人脸图像的识别率和区分性.在CASIA-webface人脸数据库进行的实验结果表明,改进后的网络模型识别率达到98. 7%,优于DCNN等算法,并将训练好的网络模型应用到YALE、PERET、LFW-A等人脸库上,相比其他方法识别率都有所提升.
Aiming at the problem that traditional neural network does not fuse the information of high and low convolution layers in the training process of face images,a fully connected convolution neural network model based on three-layer feature fusion is proposed. The algorithm combines the last three features of the original network and proposes the features. The combination of information and the last full connection layer increase the expression of shallow features,strengthen the extraction effect of deep features,and make the information extracted by the improved convolutional neural network more complete. At the same time,the loss function and the center function are weighted together to improve the recognition rate and discrimination of face images. The experimental results on CASIA-webface face database show that the improved network model recognition rate reaches 98. 7%,which is better than DCNN and other algorithms. The trained network model is applied to YALE,PERET,LFW-A and other face databases,and the recognition rate is improved compared with other methods.
作者
徐亚伟
杨会成
XU Yawei;YANG Huicheng(School of Electrical Engineering, Anhui Polytechnic University, Wuhu,Anhui 241000,China;School of Electrical Engineering, Anhui Technical College of Mechanical and Electrical Engineering, Wuhu, Anhui 241002,China)
出处
《平顶山学院学报》
2019年第2期53-58,共6页
Journal of Pingdingshan University
基金
安徽省高校自然科学研究重点项目(KJ2018A0120)
关键词
卷积神经网络
人脸识别
特征融合
中心函数
convolution neural network
face recognition
feature fusion
central function
