Deep Learning Face Recognition Algorithm For Embedded Chips

In recent years,face recognition technology has been widely applied in such industries as security,finance,entertainment and so on.Deep learning algorithm is the key to the realization of face recognition technology.However,there are still many problems in the application of deep learning face recognition technology in embedded devices.On the one hand,because there are a lot of redundant features in large-scale face training sets,the forward propagation speed of network model training is not high.On the other hand,because the structure of network model for extracting face features is complex,the algorithm runs slowly in embedded devices and the technology has not been popularized.Aiming at the above problems,this paper studies the embedded deep learning face recognition algorithm and builds the embedded face recognition system for the application scenario of automatic driving.The face data has a lot of redundant information,and it is difficult for the traditional network model to effectively extract the face features with large position changes and small size in the images as well as the embedded device of the complex network model has low running speed,so three improved methods are especially proposed to achieve the face target detection,face key point detection,face pose correction and face recognition.The main work and innovation of this paper are as follows:1.Aiming at the problem that there are a lot of redundant features in face data,an improved method of data dimensionality reduction based on LLE manifold learning is proposed.Based on LLE manifold learning,cosine distance data reconstruction method is introduced to replace the original reconstruction method,so as to solve the problem that the original method cannot effectively extract the internal manifold structure and topology structure of data when the data dimension is high.The experimental results show that the manifold structure existing in the high-dimensional data can be effectively preserved in the low-dimensional space and the forward computing speed of network is accelerated after reducing the dimension of face image.2.To solve the problem that it is difficult for the traditional network model to extract face features with large position change and small size in the images effectively,a full convolution neural network is proposed,which combines resnet18 network with improved perception.The network combines the features of resnet18 network which has more layers and perception network which has large width.Based on resnet18 network,the network width is expanded horizontally by adopting the idea of perception network,and the extraction of multiple facial features with large position change and small size is realized.Experiments verify that the full convolution neural network can effectively extract the face features of images in this paper.3.Aiming at the problem that complex network model runs at a low speed in embedded devices,small convolution kernel stack is used instead of large convolution kernel to cut the network and both methods have the same convolution field.Meanwhile,the convolution complexity of the former is relatively low and Mobile Nets is introduced to separate convolution to reduce the number of parameters and computation of convolutional layer.The experimental results show that when the network model structure is complex,the model volume and computational complexity can be effectively reduced,and the speed of the algorithm in embedded devices is also improved.4.Face recognition algorithm is implemented on embedded development board and face recognition of embedded device for automatic driving is realized.The improved algorithms proposed in this paper have better performance on embedded devices.Moreover,it can effectively extract the main features of human face and reduce the number and calculation of network parameters.