The Research Of Fine-Grained Vehicle Model Recognition Method Based On Data Enhancement

Intelligent traffic system plays an important role in assisting traffic management departments to manage vehicles,yet it is vehicle model recognition that remains an extremely important task in intelligent traffic system.In practical application,vehicle models have a wide variety of types,and some models have a high degree of similarity.Besides,different shooting angles and complex exposure situations also bring challenges to the fine recognition of models.Based on the method of generative adversarial network,this paper studies the problems of insufficient data and unequal data distribution in fine vehicle model recognition,hence proposes two ways of fine vehicle identification in different situations.They are the research on fine vehicle model recognition based on ATPGGAN and the night vehicle model recognition algorithm based on self-adaptation.Experiments were carried out on different data sets,and the results show that the two methods have good recognition effects on fine vehicle model recognition.The main work of this paper is as follows:(1)Summarized some classical theories and basic models of generative adversarial network.the theory and model related to domain adaptive are expounded.The datasets of vehicle model recognition are introduced.A night vehicle model recognition data set is disclosed.(2)An AT-PGGAN fine vehicle model recognition model based on data enhancement is proposed.The model is composed of a generation network,a classification network,meanwhile,the generation network is used to expand the training data.The attention mechanism and tag re-embedding method are used to optimize the generation network,which improves the details of image generation.(3)Put forward in view of the nighttime vehicle recognition domain adaptation methods.combining the theory of Retinex with generative adversarial network and using the improved GAN,which is based on the classical theory of image decomposition,to learn the mapping relationship between different illumination images,the image at night is converted to the one at daytime for identifying.The recognition rate on the simulated night vehicle model data set and the real checkpoint night data set was improved to some extent.