Research On Fine-grained Vehicle Classification Based On Deep Learning

With the rapid development of economy,the number of urban soars significantly,and the problem of city management and environment has become more and more conspicuous.As the key part of intelligent transportation system,Visual fine-grained classification technology is beneficial for solving those problems such as vehicle theft and criminal tracking.It could also assistant our mercantile system based on big data.However,the existing vehicle recognition methods mainly adopt cascaded networks to localize and classify targets which make it far from efficient and they are not labor friendly due to complicated label workload.Multi-layer convolutional network has advantage in high-level semantic feature extracting and it is demonstrated by multitask theory that different convolution networks could be integrated into one,sharing lower convolution layers.Therefore,we improve the convolutional network to localize cars in the image and classify them into fine-grained categories simultaneously.With the help of new loss function,the training procedure is end-to-end and there is no need for part label.The main work is as follows:Firstly,we add several global feature modules into SSD to leverage more context information in our network.Experiment results show that this method could improve detection precision for cars with limited extra computing resource.Secondly,we designed simultaneous localization and fine-grained classification network(SLAFC),in which we use GSSD-Net as a localization module and we designed a multi-layer convolutional layer as a classification module,all the modules in this framework share bottom layer by multiplexing result.The network reduces the demand for computing resource compared with traditional cascaded network and improves the accuracy,it achieves 93.3% top-1 accuracy on Compcars dataset.