Research On Efficient Retrieval Methods For Large-scale Complex Vehicle Images

With the rapid development of smart cities,to acquire and deal with a large amount of vehicle image data through video surveillance is an important way to solve public safety in smart cities,and large-scale vehicle image retrieval is one of the core links for such applications as case investigations public security departments.However,due to the widespread use of road and bayonet surveillance cameras,vehicle surveillance data has exploded,and vehicle images themselves are complex and diverse for changing quickly.Thus,a big problem occurs that it is difficult to search the target image with high similarity from the massive complex vehicle images.For instance,some vehicle images are taken in low light environment,resulting in large vehicle image noise,loss of detail information and poor visual effect,which affects the retrieval effect.Due to illumination change,viewpoint,occlusion and other factors,different vehicle images have similar models and colors,resulting in a large number of vehicle images are complex and difficult to distinguish.Vehicle images are increasing rapidly,and the scale of vehicle image data is large,which makes the retrieval efficiency low.Around these problems,the main research work of this paper is as follows:1)Aiming at the low retrieval accuracy of vehicle images due to low illumination,zero reference depth curve estimation method based on deep learning is introduced to enhance low illumination vehicle images.The source data and the enhanced vehicle image data are used to train the vehicle image retrieval model,and the retrieval evaluation indexes verify that the introduced low-illumination image enhancement method improves the accuracy of vehicle image retrieval by about 2%.2)For the problem that a large number of vehicle images are complex and difficult to distinguish,this paper proposes a multi-granularity feature fusion vehicle image retrieval model(Multiple Granularities Feature Fusion Network,MGFFN)with local branches and global branches.The local branches in this model are divided the global features into three parts evenly without additional information for labeling to fuse the local information and global information features as the required image features,and jointly use the metric loss function and the classification loss function to learn together Vehicle image features to improve the discrimination of vehicle images.The experimental results show that compared with the benchmark network model Res Net-50,the MGFFN model proposed in this paper is used with the data set Ve Ri-776 for experiments.The results of m AP,Rank-1,and Rank-5evaluation indicators show that it has improved by 15.43%、0.82% and 7.76%,respectively.And dividing the Vehicle ID dataset into different test sets of Small,Medium,and Large,the results of the evaluation index Rank-1 show that they have increased by 14.34%,13.55%,and 14.23%,respectively,and the correspondingresults using of Rank-5 has been improved by 11.35%,10.91%,11.03%,respectively.3)Aiming at the low efficiency of massive vehicle image retrieval,this paper proposes an approximate nearest neighbor HNSW vehicle image retrieval method.When the amount of data is millions,the average response time of image retrieval is milliseconds.In addition,it is proposed to apply the HNSW algorithm to the distributed architecture Varch to further improve the index construction and retrieval efficiency.When the feature vector dimension of each vehicle image is 1 000 and the number is 2.5 million,the index construction time of distributed algorithm is 2.97 times faster than that of single-point HNSW algorithm,and the response time of index query is 1.52 times faster.