Research On Large-scale Vehicle Image Retrieval Based On Convolutional Neural Network

Vehicle image retrieval is one of the important tasks in the research and development of the police video big data platform.Its purpose is to quickly and efficiently find the vehicle consistent with the query image from the massive image database,and describe its driving track according to the retrieved vehicle,which provides an important basis for the detection of cases.In recent years,content-based image retrieval(CBIR)has been widely used by users,and deep learning method has become the mainstream method for CBIR to improve the accuracy of vehicle retrieval.It mainly uses the powerful representation ability of deep neural network to image data to improve the accuracy of vehicle image retrieval.Although the method based on deep learning has made great progress in the field of vehicle image retrieval,but there are still problems with the neural network needs a large amount of data to:(1)Depth to parameter learning,on the network at present public relative lack of large-scale vehicle data set,and uses the manual annotation will spend a lot of vehicle image resources.(2)When deep neural network model is used to extract features from image database,a large number of feature vectors of high dimensions will be generated,and the time of vehicle image retrieval will be increased when similarity measurement is carried out.(3)In the process of model training,the learning and updating of the parameters of the deep neural network will incur great computational overhead,resulting in low training efficiency.Aiming at the above three problems,we propose a large-scale vehicle image retrieval method based on CNN.The research content of this article is as follows:1)The existing Vehicle ID vehicle image data set is expanded using an image data enhancement method to produce a larger vehicle image data set.The multi-scale Retinex algorithm was used for data enhancement to perform dark light enhancement on the vehicle image.Purpose is by increasing the vehicle image data to improve the depth of the neural network of image feature extraction ability,and then based on the idea of transfer learning,two different convolution neural network is adopted to Vehicle ID enhanced training data set,and data through experiment contrast,two models using data enhancement method of retrieval rate increased by 1.8% and 0.6%,respectively.2)To better apply the deep neural network to vehicle image retrieval,we propose a vehicle image retrieval method based on improved VGGNet.Firstly,a hash layer is introduced into the multi-branch VGGNet network,which maps the image feature vectors of high dimensions into binary hash codes of specified length.Secondly,VGGNet was combined with PCA principal component analysis to reduce the dimensionality of the extracted high-dimensional feature vectors.Finally,the two neural networks were trained on the vehicle data set after data amplification and verified on the test set.The experiment showed that the retrieval accuracy increased by about 1.1% and the efficiency increased by about 20%.3)On single GPU computing resources limited problems caused the network model of training time is too long,consider parallel model and the characteristics of data parallel,enhancement method based on the data and the use of the network structure was designed based on Pytorch data parallel deep learning methods to improve the efficiency of the training,using multiple GPU parallel solution to solve the bottleneck of single machine training.