基于轻量级网络的实时人脸识别算法研究

为了在嵌入式和移动设备上实现高精度的实时人脸识别,对常见的网络在人脸识别方面的优缺点进行了分析,提出了一种高效的深度卷积神经网络模型Lightfacenet。在网络中结合深度可分离卷积、逐点卷积、瓶颈结构和挤压与激励结构提出了轻量化神经网络单元,使网络在保证有一定准确率的情况下有效地解决深层的神经网络带来的参数冗余和计算量大的问题,再通过改进的非线性激活函数进一步提高网络的准确性。该神经网络在保留卷积神经网络部分优点的同时也很好地平衡了网络的缺点。在同样的实验环境下,Lightfacenet网络既实现了非常高的识别精度,也在模型推理速度上达到实时的效果。在使用MS-Celeb-1M数据集训练后,该模型在LFW数据集上达到了99.50%的准确率,其效果已经可以与现在的大型卷积神经网络媲美。对于面部识别,Lightfacenet比目前最先进的移动卷积神经网络在保证准确率的情况下提高了效率。

Research on Real-Time Face Recognition Algorithm Based on Lightweight Network

In order to achieve high-precision real-time face recognition on embedded and mobile devices,the advantages and disadvantages of common networks in face recognition are analyzed,and an efficient deep convolution neural network model Lightfacenet is proposed.In the network,a lightweight neural network unit is proposed,which combines the deep separable convolution,point-by-point convolution,bottleneck structure and squeeze and excitation structure.The network can effectively solve the problem of parameter redundancy and large computation caused by the deep neural network with a certain accuracy,and then further improve the accuracy of the network through improved non-linear activation.The neural network not only retains some advantages of the convolutional neural network,but also balances the disadvantages of the network.In the same experimental environment,the Lightfacenet network not only achieves very high recognition accuracy,but also achieves real-time effect in the speed of model reasoning.After trained on MS-Celeb-1 M data set,this model achieves 99.50%accuracy on LFW,and its effect is comparable to the advanced large deep convolutional neural network.For face recognition,Lightfacenet improves efficiency while ensuring accuracy compared to the most advanced mobile convolutional neural networks.