Vehicle Feature Learning And Vehicle Recognition

The rapid development of the automobile industry has brought great convenience to our lives,but also brought many problems,such as traffic jam,vehicle mugging,crime with vehicle,etc.With the existing technology of automatically identifying vehicle models,the above problems can be quick to be responded and effectively resolved.How to reliably extract and describe the features of vehicle images are the key topics of vehicle feature learning and vehicle identification.What's more,the establishment of a stable vehicle identification model and real-time recognition of large-scale vehicle images are also vital.In this dissertation,the essential features of vehicle images were analyzed and discussed,and the corresponding vehicle identification framework was constructed.The main contents are as follows:1.I analyzed and discussed the performance of the traditional features of vehicle images,and constructed the vehicle identification framework based on the traditional features.SIFT,HOG and LBP as the widely used hand-craft features are firstly reviewed.Then I introduced the mid-level features BOF,FV and VLAD of vehicle images.Experiments show that the discrimination of conventional features is insufficient for the vehicle identification.2.The deep features of the vehicle image were thoroughly investigated,and the vehicle recognition framework based on the fusion deep features was established.In order to solve the disadvantage of the existing deep network,the Inception module of GoogLe Net was further decomposed,and I tried to merge the Inception module and ResNet to build the framework of vehicle recognition.The experimental results demonstrate the recognition rate of the proposed method is improved for vehicle identification.3.I analyzed the dense deep features of the vehicle image,and the natural expression of features for vehicle image is explored.For the difficulty in training the deep network,the ResNet with residual connection was used to learn the dense deep features.Moreover,the center loss function was combined into the vehicle identification framework to improve the discrimination of the dense deep feature.The experimental results show that the dense deep feature is more compact,the number of parameters isreduced,and the recognition rate is further improved.