Vehicle Recognition Algorithm Based On Deep Learning

Vehicle recognition is an important part of intelligent transportation.It plays an important role in various fields such as high-speed toll collection system,driverless vehicle and so on.As a necessary transport for daily life,there are various kinds of vehicle types,so it is meaningful to make research about the vehicle recognition.In order to solve low recognition accuracy caused by the lack of training data or the rec-ognition performance constraint of single convolutional network model.A vehicle recognition algorithm based on data augmentation and multi-model ensemble is pro-posed.In order to avoid the influence of complex background on Fine-Grained recog-nition,as well as the study of the global information and local information of the tar-get objects extracted from the convolutional neural network for fine-grained tasks.A vehicle recognition method based on multi-scale region feature is proposed.The main research contents are as follows:(1)Vehicle recognition algorithm based on data augmentation and multi-model ensemble firstly designs a variety of data augmentation methods to increase the num-ber of pictures in CompCars dataset,including mirroring,rotation,gaussian noise,and color enhancement.Then 3 differentiated models,ZF,VGG 16 and ResNet,are trained using the constructed differentiated dataset by different data sampling.A ensemble learning method is used to integrate multi-model's recognition results.The experi-mental results show the fine-grained recognition of the differentiated convolution network trained on the different datasets generated by the different data augmentation method.The experiment also shows the recognition results of multi-model ensemble.The final recognition accuracy of multi-model ensemble is 95.0%.Compared with GoogleNet recognition algorithm,the recognition accuracy is increased 3.8%.The results verify the effectives of proposed algorithm.(2)Vehicle recognition algorithm based on Multi-scale region feature firstly uses FASTER-RCNN framework to train three convolution models to locate multi-scale object regions.Then the bounding box constraint and Helen constraint are applied to improve the location accuracy of the detected object.Finally,the extracted mul-ti-scaled region features are combined to train a SVM classifier for fine-grained rec-ognition.The proposed method is tested in CompCars vehicle datasets.The experi-mental results show that the accuracy of recognition in CompCars is 93.5%.It in-creases 8.3 percents than the method without multi-scale region features.Compared with GoogleNet recognition algorithm,the result improves by 2.3 percents.