Research And Implementation Of Face Recognition Based On Deep Learning

Face recognition technology has always been one of the hotspots in the field of computer vision.At the same time,with the rise of artificial intelligence technology,face recognition is also a hotspot for deep learning algorithms.It has wide applications in security,transportation,medical and many other fields,prospect.Face recognition technology includes face detection technology and face feature extraction technology.The face recognition technology of traditional artificial feature extraction method has been affected by many aspects for a long time,making the accuracy of face recognition technology always low and unable to be practical.Apply to people's production and life.Thanks to the development of deep learning,deep learning is better and better for extracting image features.It can extract deep and essential features of images,which can fully represent faces and have strong generalization ability.This makes the use of deep learning to solve the face recognition problem that traditional methods can't solve has become the focus of academic and business circles.At present,the face recognition algorithm based on deep learning is still in the process of development.There are some problems:1)How to design a face recognition network with high accuracy and low complexity,and compress the model while ensuring the recognition accuracy.The amount of parameters makes the hardware resources occupied by the face learning model based on deep learning as small as possible;2)how to design a loss function suitable for face recognition,so that the face recognition network can be trained more quickly,while the network Better performance.In this paper,the current situation and existing problems in the field of face recognition research are explored,and a feasible face recognition method based on deep learning is proposed.The main contributions include:1)Convolution analysis in this paper The development of neural network and the principle of each convolutional neural network model,deeply analyzed the network characteristics of lightweight network SqueezeNet,combined with the idea ofFire Module in SqueezeNet network,designed a lightweight face recognition network,the network face The recognition accuracy is better,and the parameter quantity is much lower than other deep neural networks.2)For the problem caused by the single loss function,the inter-class difference and intra-class difference of face recognition problem are analyzed and designed.A mixed loss function for weighted fusion Softmax loss,central loss,and verification loss.In this paper,the face recognition network based on SqueezeNet network combined with the mixed loss function of Softmax,center loss and verification loss get the accuracy of 99.23%on the LFW verification set,and the parameter quantity is only 1/5 of the VGG-16 network.On this basis,an online face recognition system is implemented by using the trained model and Tensorflow neural network framework.The results show that the model is effective and can be implemented and applied well-Finally,the paper summarizes the work obtained in the process of using the deep learning method to solve the face recognition problem,plans the next step,prepares to continue to study the face recognition method based on deep learning,and will reduce the lightweight neural network SqueezeNet Deploy to the terminal device.