Research On Multi-scene Model Recognition Based On Deep Learning

As an inevitable choice for the development of smart cities,intelligent transportation systems have shown unique advantages and potentials in vehicle control and road management.In the intelligent transportation system,vehicle detection and identification is an important part of the system.The classification of vehicle models is a key task.It can effectively reduce traffic load,improve the operational efficiency of transportation facilities and the utilization of urban resources,and provide people with daily travel.Basic security.At the same time,in the field of image processing and pattern recognition,deep learning based on computer vision has achieved rapid development and wide application.Therefore,this paper applies the convolutional neural network idea in deep learning to the research of vehicle identification problem in multiple scenes,in order to obtain higher efficiency and more complete data information,greatly promote the optimization and upgrading of vehicle classification technology,and realize information.Intelligent processing to improve traffic efficiency and solve a series of problems in traffic management and road monitoring.In order to solve the problems of single feature identification,low recognition accuracy and low efficiency of traditional vehicle identification methods,the convolutional neural network is introduced into the target recognition problem,and a vehicle identification method based on improved convolutional neural network is proposed,which is clear and efficient.The generalization ability to complete the feature learning of the model image.In-depth study of the vehicle classification task model for the following scene images was carried out:1.Aiming at the vehicle identification method under the high-speed road video surveillance scene,the paper mainly explores the framework design and internal parameter optimization of the convolutional neural network model,and constructs a multi-level image classification framework based on the cyclic neural network.The traditional model pooling method and parameter tuning method are designed based on the improved Alex Net network model to broadenthe horizontal structure of the network model to achieve the diversity of feature extraction during the training process,through the BIT-vehicle dataset.A series of comparative experiments show that the proposed model is effective in solving the coarse-grain classification problem of the vehicle.Compared with the improved convolutional neural network method,the accuracy and efficiency of the optimized model recognition are significantly improved.2.The refined classification method for static vehicle images,the double-branch convolutional neural network model is designed around the multi-task learning method that shares most of the neural network layers,and the complex vehicle identification problem is decomposed into simple and independent.The two sub-problems of vehicle brand and type,through the establishment of shared underlying network structure to achieve joint training of various related tasks,improve the generalization effect of the model,and then merge the branch output to get the final result,to achieve the refined classification task requirements.The comparison experiments on the Stanford-16 dataset show that multi-task learning has certain advantages over the single-task learning method in the refined classification performance of the vehicle,showing a high recognition accuracy,reflecting the high-quality prediction ability of the model.