基于轻量级卷积神经网络人脸识别算法的研究与应用

Research and application of face recognition algorithmbased on lightweight CNN

针对多数卷积神经网络(CNN)人脸识别算法在追求高精度的同时,加大网络层数,造成网络参数过多、计算量过大、训练速度缓慢且对设备要求高,增加了成本的问题,提出了一种改进型轻量级卷积神经网络MobileNet的人脸识别算法。首先,将MobileNet中的SoftMax层为L-SoftMax层,避免了过度拟合,实现更好的分类效果。其次,将改进的MobileNet和区域生成网络(RPN)融合,并在Jetson Nano小型设备上进行训练。实验表明:所提算法与传统的卷积神经网络人脸识别算法相比,在LFW人脸数据库和自建的小型数据库上训练测试,模型的参数量减少了88%,识别准确率与原MobileNet相比增加了0.2%,达到了97.54%。运行速度较原MobileNet网络提高了21.3%。

Aiming at the problem that many convolutional neural network(CNN)face recognition algorithms pursue high precision,at the same time,increase the number of network layers,resulting in excessive network parameters,too much calculation,slow training speed and high equipment requirements,which increases the cost,an improved lightweight CNN face recognition algorithm is proposed.It is mainly implemented by Mobilenet.Firstly,the SoftMax layer in MobileNet is replaced by L-SoftMax layer.L-SoftMax avoids over fitting and has better classification effect.Secondly,the MobileNet will be integrated with the region proposal network(RPN).Experimental results show that,compared with the traditional CNN face recognition algorithm,the proposed algorithm can reduce the number of parameters by 88%and increase the recognition accuracy rate by 0.2%compared with the original MobileNet algorithm,reaching 97.54%after training and testing on LFW face database and its own small database.Compared with the original MobileNet,the running speed is increased by 21.3%.