面向边缘设备的端到端人脸检测网络

End⁃to⁃end face detection network for edge devices

针对常用的人脸检测算法模型训练难、计算量大,导致算法难以部署到边缘端设备的问题,提出一种基于轻量化SSD的人脸检测算法。首先,在数据集中加入部分模糊和有遮挡的含噪声样本,以此来增强训练数据集;然后,引入深度可分离思想对SSD目标检测网络进行轻量化操作,通过增强数据对网络进行训练,获得用于人脸检测的轻量级神经网络;最后,将实验模型在FDDB和Yale-face数据集上进行实验,在FDDB数据集上正确检测率达到了97.03%,在Yale-face上的检测率则达到了87.42%。实验结果表明,所提算法在保持精度的同时具有较低的模型复杂度和计算量。

In view of the facts that it is difficult to train common face detection algorithm models and difficult to deploy the algorithm to edge devices due to the large amount of computation,a face detection algorithm based on lightweight SSD is proposed.Firstly,the training data set is enhanced by adding some fuzzy and obscured noise samples to the data set.Then,the idea of depth separability is introduced to conduct lightweight operation on the SSD target detection network,and the network is trained by enhancing the data to obtain a lightweight neural network for face detection.Finally,the experimental model was tested on the FDDB and yale⁃face data sets.The results show that the correct detection rate on the FDDB data set reaches 97.03%,and the detection rate on yale⁃face reaches 87.42%.They prove that the proposed algorithm has lower model complexity and less computated quantity while maintaining the accuracy.