Research On Vehicle Logo Localization And Recognition Based On Feature Learning

Intelligent transportation system(ITS)plays an important role for the increasingly serious traffic problems happened.Vehicles are the principal part of ITS so that the vehicle recognition is the key issue that need to be solved in ITS.Vehicle logo is the key identifier of a vehicle' brand,so we can obtain the brand information using the vehicle logo recognition technology.Vehicle brand information can help us identify whether the information comes from the license plate recognition is accurate or not,thus can avoid the problem caused by false information come from changed or modified license plates,and reduce unnecessary loss.The vehicle logo recognition task includes two parts,vehicle logo localization and classification.The traditional vehicle logo localization and classification methods are mainly based on hand-crafted descriptors which need prior knowledge and are difficult to adapt the real application environments.Therefore,focusing on the research of feature-learning based methods,this paper proposes a vehicle logo localization method based on improved Yolov3 algorithm and a vehicle logo recognition method based on shallow feature learning.These methods can automatically learn feature information,thus avoid the shortcomings of handcrafted descriptors.Moreover,our experimental results prove that it has good performance than many other state-of-the-art methods.The main work can be concluded as follows:(1)The existing vehicle logo location and classification algorithm analysis: the advantages and shortcomings of both the traditional hand-crafted methods and deep learning based methods are analyzed.Some classical algorithms used for object detection and classification,and some popular datasets are introduced.(2)A vehicle logo localization method based on improved Yolov3 is proposed.The method is called VLD-Net.First,we introduce the series method of Yolo,and analyze the advantages and disadvantages of Yolov3.Then,we improve the Yolov3 algorithm and make it more suitable for the task of vehicle logo location.(3)A vehicle logo classification method based on shallow feature learning called MLP-Net is proposed.Through optimization of the objective function,the feature parameters are learned,and then feature information are extracted and feature vectors are generated.In this way,it avoids the shortcoming of handcrafted descriptors.Moreover,MLP-Net contains less feature parameters.Therefore,it can converge with small training samples and take much less time.Compared with deep-learning based methods,the MLP-Net takes much less time but can still have very high recognition accuracy.