Research On Vehicle Face Recognition Based On Deep Learning

With the development of society,vehicles have become one of the indispensable means of transportation for human beings.While vehicles brought convenience to human beings,they also led to a series of problems.In addition to the problems of traffic congestion and environmental pollution,there were many traffic accidents,which threatened the safety of human life and property,so the management of vehicles is very important.Adding license plate to a vehicle is one of the means to manage the vehicle,which is equivalent to the "identity card" of the vehicle.When the driver breaks the law,the managers can find the vehicle throught license plate number to punish the driver accordingly.But now the technology of license plate recognition is relatively mature and the license plate information is easy to obtain.Many drivers use false license plates or hide and dirty the license plate in order to avoid punishment.They drive on the road recklessly,which causes great difficulties to the supervision of vehicles.Aiming at these problems,vehicle face recognition technology emerges as the times require.With the development of deep learning in recent years,vehicle face recognition is developing towards a faster and more accurate direction.In this paper,a method of vehicle face recognition based on deep learning is studied.The feature extraction of vehicle face images is carried out by convolution neural network,and the classification and recognition of images are carried out by Faster R-CNN.The main research work of this paper is as follows:1)A car face image database is collected and calibrated.Aiming at the problem that the original ZFNet convolutional neural network has a low level of feature extraction for the car face,an improved ZFNet convolutional neural network is used to extract deeper features of the car face,so as to improve the classification recognition rate of the car face;On their own when the car is face feature extraction,face classification recognition rate low,was designed and implemented based on improved ZFNet car components face feature extraction method,the method based on car face component location selection face of five components: license plate,car logo,car grille,car lights and car windwhile,components on a single car face feature extraction;Affine transformation is introduced to achieve the alignment of the car face,and the complete normalized car face is obtained as the input of the subsequent network.2)After ZFNet convolution neural network extracts vehicle face features,Faster R-CNN is used to compare the eigenvalues of the extracted vehicle face features.After defining the feature category,the new vehicle face image can be classified into the category after entering the network,so as to achieve the purpose of vehicle face recognition.The RPN network in Faster R-CNN is studied.In order to improve the classification and recognition efficiency of Faster R-CNN,IOU threshold and loss function of RPN are modified to achieve better classification and recognition effect.3)In order to further improve the accuracy of classification and recognition of this data set,a multi-face component recognition method based on decision tree is adopted to further ensure the accuracy of the results of recognition data through decision tree.The combination of recognition results of face components is used to judge the whole face,which enhances the association between components and improves the accuracy of face recognition.