Research On Vehicle Recognition And Vehicle Retrieval Based On Deep Learning

With the rapid development of our country's economy and society,the number of cars in various regions also increases rapidly.In the face of all kinds of vehicles on the road,an effective algorithm is designed to enable it to accurately identify the brand and model of the vehicle,and to retrieve the vehicle image of the same model from the vehicle database,which is of great significance to the development of intelligent traffic.In the past,the recognition of vehicle image is mostly coarse-grained general vehicle recognition,and there are few researches on fine-grained vehicle retrieval.Therefore,based on the deep learning technology and combined with the target detection algorithm,the corresponding vehicle recognition and vehicle retrieval methods are given in this thesis.The main research contents and methods of this thesis include the following three aspects:(1)This thesis summarizes the basic principles of existing deep learning models,the target detection methods based on deep learning,and the target detection algorithms based on region and regression.The development trend of deep learning in vehicle recognition and vehicle retrieval is also discussed.(2)An improved Faster R-CNN fine-grained vehicle recognition algorithm is proposed.This algorithm improves Faster R-CNN based on deep learning target detection algorithm to increase vehicle recognition accuracy and efficiency.The main contributions of this thesis are listed as follows: 1)multi-scale features containing rich semantic information were extracted by combining ResNe Xt and FPN;2)improve the proportion scale of anchor frame generated by RPN to increase the vehicle identification efficiency;3)reduce the missed detection rate through SoftNM and RoIAlign;4)join the online hard sample mining mechanism to solve the problem of unbalanced positive and negative samples;5)use Dropout to reduce network generalization error and avoid overfitting.Finally,the experiments are carried out in CompCars and Stanford Cars,a vehicle image database of different sizes,and the effectiveness of the proposed method is verified by the comparisons of the algorithms.(3)A two-stage fine-grained vehicle image retrieval algorithm based on deep learning is proposed.First of all,the algorithm selects features that contain valid information,globalfeature descriptors are generated by generalized mean pooling,and then the initial retrieval results are obtained by Euclidean distance method.In the second stage,the classification score and position coordinates of the target region are predicted by Faster R-CNN,and the same image as the query category is found in the initial search result.Combined with the query expansion,the Euclidean distance calculation is carried out again to retrieve the final similar image.The experiment was carried out on the vehicle database CompCars and Stanford Cars,the effectiveness of the proposed algorithm was proved.