Application Of Deep Learning In Vehicle And Face Recognition

Nowadays,intelligent transportation systems play an increasingly important role.Accurate collection and processing of information through monitoring systems is the basis for the effective operation of intelligent transportation systems.Vehicle recognition and face recognition are information processing technologies that are commonly used in intelligent transportation systems,and play an important role in practical applications.With the development and the increasing progress of deep learning technology,more and more technical problems have been solved,and the increasing traditional technologies have been replaced by deep learning.It has important research value and application prospect that convolutional neural networks are applied to vehicle recognition and face recognition and are ameliorated for practical problems so that the accuracy of information collection and processing in ITS are improved.The convolutional neural networks in deep learning are studied and analyzed in this paper.And they are applied in vehicle recognition and face recognition by improving the network structure.The main tasks completed are as follows:(1)Classification and recognition of multi-scale vehicle targets are researched.A relatively complete dataset which contains vehicle images of six different target proportions is built.Multiple convolutional neural network structures are constructed and tested by image samples of different target proportions so that influence of different target proportions on the accuracy of convolutional neural networks is studied.And then,the network structure with stable performance and high accuracy is gotten through comparing and analyzing experimental results.The classification ability of the network is tested by image samples with multiple target proportions.The experimental results show that the network structure can effectively reduce the influence of the target proportions in the multi-proportional vehicle target classification and recognition,and the accuracy is always higher than 96.33% and maximum fluctuation not exceeding 1%.(2)Multi-scale face recognition is studied.For the fixed input size of convolutional neural network,a multi-scale face recognition network structure combining with the bag of words model is proposed.The two commonly used face datasets,ORL and Yale B,are scaled to 5 sizes to test network performance.The number of optimal cluster centers is obtained after multiple experiments.And then recognition ability of the network for different sizes is tested.Finally,the proposed arithmetic is compared with GAP,GMP and SPP.The experimental results show that the proposed arithmetic gets a higher accuracy in multi-scale face recognition.In the exexperiment of Yale B,the accuracy of size of 192×192 is 97.50%.(3)Driver face recognition is studied.For less training samples in driver face recognition,the Siamese network is introduced to solve the problem of one shot learning and is applied to recognize driver face.The drivers’ face images and the face images in life are collected,and the drivers’ one-sample dataset are created.The Siamese network structure is constructed.After multiple iterations of training and testing,the effectiveness of the network in driver face recognition is verified and good results are obtained.The accuracy of the model for single test samples of 60 categories of faces can reach 98.83%.