基于改进的SqueezeNet的人脸识别

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。

Face Recognition Based on Improved SqueezeNet

The traditional machine learning face recognition method is easily affected by conditions such as angle and illumination, and is improved after the convolutional neural network is proposed. In view of the current convolutional neural network to improve the recognition rate, the model parameters are neglected, resulting in too large a parameter size to run on hardware with small memory. An improved model based on SqueezeNet is proposed to identify the face. The SqueezeNet model uses small convolution check kernel feature extraction, which is beneficial to reduce the model parameter input. However, small convolution kernels tend to over-fitting on complex training sets. In order to prevent over-fitting, the improved SqueezeNet model introduces the first pooling layer and the last pooling layer into the next layer of convolutional layer fusion, which can extract more feature information and optimize the loss of feature information. For the improvement of the classification function Softmax, the L2 norm constraint method is used to constrain the features of the last layer into a spherical surface, which is easy to converge and prevent over-fitting. The experiment validated the validity on the CASIA-webface and ORL face databases.