The Research Of Vehicle Logo Recognition Method Based On Feature Learning

As an important part of the intelligent transportation system,vehicle logo recognition(VLR)has a wide range of application prospects in the field of vehicle and traffic management.Moreover,VLR also provides information clue for vehicle identification,thus it has important research value.Most of the existing VLR methods are based on hand-crafted features,which usually require strong priors to design them by hand.Their performance has poor robustness in complex environments.Compared with the hand-crafted features,the VLR methods based on feature learning has a great advantage in solving the VLR problem in complex environments.Therefore,this thesis focuses on the feature learning method of vehicle logo recognition.The main work of this dissertation listed as follows:(1)Existing vehicle logo recognition methods are analyzed:existing methods are concluded,including vehicle logo recognition algorithms based on traditional features and deep learning,and the characteristics of each algorithm are analyzed.(2)A coarse-to-fine vehicle logo detection method is proposed:by combining the position of the vehicle logo with the license plate and vehicle face,a coarse-to-fine vehicle logo detection method is proposed.Based on the license plate and the vehicle frontal face,the rough target region of the vehicle logo is firstly detected.Then the features of the logo area and non-logo area are extracted.A classifier model is trained to achieve the fine detection of vehicle logo.(3)A vehicle logo recognition method based on feature learning is proposed:by combining the local sampling method in traditional features and the network structure in deep learning,we propose a feature extraction method based on learning.By converting the extracted pixel gradients to a low-dimensional binary feature vector,the feature we extract become more compact and robust.