Research And Implementation Of Face Recognition Algorithm Based On Jetson TX1

With the rapid development of biometric identification technology,more and more places are beginning to use biometrics.From the launch of Apple's Face ID,it can be seen that biometric recognition changes from fingerprint recognition to face recognition.The most important advantage of face recognition is the Non-contact collection,which will not make users feel violated,but also accord with the human identification habit.The classic face recognition framework divides the face recognition system into three parts: face detection,face alignment and face verification.Face recognition technology has a long history of development,but it has poor robustness to changes caused by non-ideal acquisition environments or non-ideal acquisition environments due to non-ideal lighting and posture.The traditional face recognition method is based on the characteristics of human manual design and its accuracy is about 85%(human recognition accuracy rate is 94.27%).The traditional method runs too long and needs nearly 1s,and it is difficult to reach the practical level.With the development of high performance graphics cards,convolutional neural networks have been widely used in the field of computer vision,and have significantly improved the existing face recognition technology.However,high performance graphics cards are large in size,and the operating power consumption reaches 300 W and the price of tens of thousands of them,making it difficult to deploy in many lightweight environments.This article summarizes existing methods for face recognition.The traditional method has low recognition accuracy and long running time.Face recognition methods based on high performance graphics cards and convolutional neural networks have disadvantages such as high power consumption,high cost,and large size.So this article proposes multitask convolutional neural networks for face detection,stacked hourglass structures for face alignment,and mobile networks that fuse center loss Softmax functions for face verification.Finally,this article uses Jetson TX1,a credit card-sized board,to implement the above-mentioned fast and accurate offline face recognition system.First of all,the classic traditional face recognition algorithm is deeply studied.The recognition results in the natural environment are compared and their respective advantages and disadvantages are also compared.It is concluded that the traditional algorithm is not applicable in the natural scene demanded by this article.So a facial recognition algorithm based on Jetson TX1 and deep learning is proposed.Secondly,this article deeply researches the face detection network,the face network and the face verification network.In the aspect of face detection,the cascaded convolutional neural network is improved and a multitask convolutional neural network is proposed.In the aspect of face alignment,the face feature localization network of the stacked hourglass structure is proposed based on the very effective heat map regression-based full convolutional neural network in the human pose estimation.In face verification,this article uses depth-separable convolution to reduce network parameters without degrading recognition accuracy.For the first time,this article proposes to use lightweight mobile network to extract features and introduce center loss function to improve the accuracy of the verification network.Finally,this paper briefly introduces the Jetson TX1 performance of the algorithm loading platform.In order to reflect the recognition accuracy and running time of this algorithm.This article tests a large number of pictures in various natural scenes,including strong light,dark,side face,partial occlusion,face tilt,exaggerated facial expressions.This article summarizes the algorithms and existing algorithms from the aspects of accuracy and running time,and analyzes their advantages and disadvantages.Compared to existing face recognition algorithms,the algorithm in this paper makes the best trade-off between accuracy and speed,and can run on the Jetson TX1 with only a credit card size.Faces of various scenes in this article can be well recognized.