基于SE-ResNet模型的多角度人脸识别系统设计

Design of multi angle face recognition system based on se RESNET model

由于在现实中人脸识别系统采集的图像大都是侧脸图像,严重影响了人脸识别的准确性。提出一种以SE-ResNet模型为基础的多角度人脸识别系统。该系统将SE-ResNet网络用于人脸特征提取器,通过在ResNet基础上嵌入SE模块,根据网络获取的特征重新标记,提高有效特征信息占比的同时,减少无效特征的数据信息。使用损失函数ArcFace进行图像训练,将角边距引入到余弦角度,以加强角度空间对于人脸特征限制条件,进而增加不同类型人脸的差异性与相同类型特征间紧凑性。通过实验结果分析可知:与基于ResNet的人脸识别方法相比,本文研究的多角度人脸识别方法准确率提高了1.9%。

In reality,most of the images collected by face recognition system are side face images,which will seriously affect the accuracy of face recognition.A multi degree face recognition system based on se RESNET model is proposed.In this system,Se RESNET network is used for face feature extractor.By embedding se module on the basis of RESNET and re marking according to the features acquired by the network,the proportion of effective feature information is increased as much as possible and the data information of invalid features is reduced.At the same time,the loss function arcface is used for image training,and the corner margin is introduced into cosine angle to strengthen the restriction of angle space for face features,and then increase the difference between different types of faces and the compactness between the same type of features.The experimental results show that:compared with RESNET based face recognition method,the accuracy of the proposed multi angle face recognition method is improved by 1.9%.