Vehicle Part Recognition Based On Co-occurrence Relationships

With the continuous development of the artificial intelligence technology,Intelligent Transportation Systems have been more and more widely used in the transportation industry.As an important technology in Intelligent Transportation Systems,vehicle part recognition has attracted the attention of more and more researchers.This technology aims to determine the subcategories of each vehicle part by automated means,which can be applied to scenes where the license plate number cannot be identified or the camera needs to automatically recognize a certain model of the vehicle.At present,existing algorithms consider to recognize each category as independent classification tasks,which ignore the potential correlation between different recognition tasks.By analyzing the distribution law of the datasets,we find that there is a certain relationship between the categories of different vehicle parts,which can be referred as the "co-occurrence relationships".It can be explained as the likelihood that a subcategory of one vehicle part can determine the subcategory of another vehicle part on the same vehicle.In order to improve recognition precisions,we propose three vehicle part recognition networks based on co-occurrence relationships.First of all,a method to evaluate the strength of the co-occurrence relationship is given---mutual information.By calculating the mutual information,it can be found that different vehicle part combinations own different strength of co-occurrence relationships.Secondly,the paper presents a vehicle part recognition network based on Siamese structure.The network references Siamese network and uses the deep residual network as the basic network.It uses a network of the same structure and weight sharing to make use of co-occurrence relationships.In addition,in order to complete the classification task,the paper modifies the similarity measure loss function at the end of the Siamese network to a cross entropy loss function.Experiments show that the vehicle part recognition network based on Siamese structure can achieve a better recognition effect on most vehicle parts than the traditional recognition network for a single vehicle part.Aiming at the problem that the recognition performance of the vehicle part recognition network based on Siamese structure have got less promotion,the paper proposes a vehicle part recognition network based on channel merging.The network uses channel merging to integrate the co-occurrence relationship information carried by the input vehicle part images.This approach not only allows each recognition task to share weights,but also allows the network to extract features simultaneously from multiple vehicle part of the input.By channel integration of the input images,the network can learn and utilize co-occurrence relationships better.Experiments show that the vehicle part recognition network based on channel merging can improve the recognition precision to a greater extent than the vehicle part recognition network based on Siamese structure.Although the vehicle part recognition network based on channel merging allows the network to extract features from multiple vehicle parts at the same time,this method also causes the information of each vehicle part to be prematurely integrated with the information of other vehicle parts,thereby losing the interclass difference information of each vehicle part itself.In order to solve this problem,the paper proposes a vehicle part recognition network based on semantic feature merging.The network first uses two independent sub-networks to extract the semantic information of the two types of input vehicle parts.Then it learns co-occurrence relationships through a network of weight sharing.This two-stage structure not only allows the network to focus on the co-occurrence relationships between vehicle parts,but also considers the interclass differences of each type of vehicle parts.Experiments show that the network has better recognition performance than vehicle part recognition networks based on channel merging.