Fine-Grained Vehicle Recognition Based On Deep Learning

Fine-grained vehicle recognition is the core of Intelligent Transportation System.It has a wide range of applications in smart parking lots,highway intelligent toll collection,vehicle intelligent management,etc.In particular,the model for fine-grained vehicle recognition can effectively assist vehicle retrieval and comparison,and help to crack down on vehicle violations such as fake plate,run accident and so on,and promote the maintenance of public security.Therefore,the research of finegrained vehicle recognition has wide application value.In this thesis,the fine-grained vehicle recognition is studied,which is divided into two types: finegrained vehicle recognition under limited angle and fine-grained vehicle recognition under multi-angles.The main application scenarios of fine-grained vehicle recognition under limited angle are the monitoring system of the entrance in expressway,national highway and provincial highway.Only frontal images of vehicle can be adopted.And there are serious data imbalances among different vehicle types.The application of fine-grained vehicle recognition under multi-angle is more extensive.It is hard to obtain effective and stable visual features for vehicle under multi-angles,because of complex background and changeable angles.In view of the problems above,the main work of this thesis is as follows:Firstly,in order to deal with the problem of class-imbalance,we adopt multi-strategies.In the stage of data pre-processing,data augmentation is used to increase the number of samples of the minority class.In the stage of randomly sampling of training data,the proportion of the majority class is controlled in the continuous batch data.In the stage of training classifier,CS-SVM is used.Secondly,in order to solve the problem that only frontal images of vehicle can be obtained under limited angle,we propose a part-based model for fine-grained vehicle recognition in a weakly unsupervised manner.This thesis also provides a part location method that locates the discriminative parts based on saliency maps which can be easily obtained by a single backpropagation pass.The advantage of the method is that the resolution of saliency maps is the same as the resolution of input images.Thus,we can locate discriminative parts efficiently and accurately.Then,aiming at the problem of complex and changeable background of vehicle under multi-angles.The object of vehicle in the image is detected by the SSD based on ResNet-50 to reduce the interference of complex background on fine-grained vehicle recognition.Lastly,aiming at the problem that it is difficult to obtain effective and stable visual features for vehicle under multi-angles,two schemes are proposed: fine-grained vehicle recognition based on multiangles feature fusion and fine-grained vehicle recognition under multi-angles based on MS-B-CNN,and the experimental results of the schemes were compared and analyzed.