Abstract: Vehicle identification, especially fine-grained vehicle identification, is an important part of
modern intelligent transportation system. Aiming at the problem that it is difficult to effectively
recognize fine-grained vehicle using traditional vehicle identification methods, we take three classic
deep convolutional neural networks (such as AlexNet, GoogleNet and ResNet) as the basic networks,
and introduce the classification of vehicle types as the auxiliary task, together with fine-grained
vehicle identification task to constitute a multitask joint-learning model. By training and evaluating
our model on a public data set which contains 281 vehicle types, we have demonstrated the excellent
performance of this model in fine-grained vehicle identification task with no need of annotations
about vehicle parts’ location and additional 3D information. Besides, with the introduction of
multitask learning strategy, the performance of this model can be improved, compared with that of
any single-task learning model. Our model is simple and efficient, and can basically meet the demand
of practical applications.

Key words: intelligent transportation, fine-grained vehicle identification, deep learning, neural
networks