基于改进的SqueezeNet的人脸识别  

Face Recognition Based on Improved SqueezeNet

作　　者：吴军[1] 邱阳 卢忠亮[1] WU Jun;QIU Yang;LU Zhong-liang(Jiangxi University of Science and Technology,College of Information Engineering,Ganzhou 341000,China)

机构地区：[1]江西理工大学信息工程学院,赣州341000

出　　处：《科学技术与工程》2019年第11期218-223,共6页Science Technology and Engineering

基　　金：国家自然科学基金青年项目(61701203);江西省教育厅科技项目(GJJ150642)资助

摘　　要：针对众多基于卷积神经网络的人脸识别技术在追求提高人脸识别率上,忽视了网络模型输入参数,导致模型输入参数多、训练时间长和无法在内存小的硬件上运行等问题,提出一种基于改进的Squeeze Net的人脸识别模型。改进的Squeeze Net模型保留了原网络模型中的小卷积核去提取图片特征,还采用首尾池化层分别引入对应的后续卷积层进行特征融合,提取细微的人脸纹理特征来稳定模型收敛性,防止小的卷积核在复杂的人脸训练集上产生过拟合。针对分类函数Softmax的改进,采用L2范数约束的方法,将最后一层的特征约束在一个球面内,减少相同特征间距,提高网络收敛能力。通过两种改进后的Squeeze Net模型在与其他的先进模型对比,在不降低人脸识别准确率的前提下,具有输入参数少、模型易于收敛和能够运行在内存小的硬件设备的优势。结果在CASIA-WebFace和ORL人脸库上得到了有效性的实验验证。For the pursuit of improving face recognition rate, many face recognition technologies based on convolutional neural networks are neglecting the parameters of the input network model, resulting in many models with many input parameters, long training time and inability to run on hardware with small memory. A face recognition model based on the improved SqueezeNet. The improved SqueezeNet model retains the small convolution kernel in the original network model to extract image features. It also uses the first and last pooling layers to introduce corresponding subsequent convolutional layer fusions to extract subtle facial texture features and stabilize the model convergence. Preventing small convolution kernels produce over-fitting on complex face training sets. For the improvement of the classification function Softmax, the L2 norm constraint method is used to constrain the features of the last layer in a spherical plane, reduce the same feature distance, and improve the network convergence ability. Compared with other advanced models, the two improved SqueezeNet models have the advantages of less input parameters , shorter training time, and the ability to run hardware devices with small memory without reducing the accuracy of face recognition. The results were validated experimentally on CASIA-WebFace and ORL face database.

关 键 词：卷积神经网络 人脸识别 特征融合 分类函数 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]