Research On Vehicle Classification Based On Improved CapsNet

With the development of big data technology and urban informatization,intelligent transportation systems have become an important part of smart cities.Intelligent transportation systems can improve road traffic efficiency and reduce the occurrence of traffic accidents.Vehicle classification is an important part of intelligent transportation systems.The intelligent transportation system provides data support,which can improve people’s travel efficiency,provide technical support for combating crimes such as accidents and escapes.Caps Net show excellent performance in image processing,which brings more new solutions to vehicle classification.This paper summarizes and analyzes the Caps Net.Taking advantage of its ability to efficiently extract image features,pre-convolutional Caps Net and Caps Net infused gradient convolution features were proposed to complete the vehicle models classification in the traffic surveillance scenario:1.Aiming at the problem that the recognition accuracy of vehicle classification in existing road traffic is not high,Caps Net can be used to extract the advantages of more features,and a pre-convolution Caps Net model was proposed.Based on the Caps Net,the model adds a new multi-layer convolution to extract shallow features,performs nonlinear activation through the Leaky Re LU function,uses the Adam optimizer to optimize,and makes adaptive improvements to the capsule network part.Experiments were conducted on the BIT-Vehicle public dataset,and the comprehensive evaluation index of pre-convolution Caps Net identification reached 97.91%,and the accuracy reached 98.33%,which is superior to the identification of other network structures such as traditional convolutional neural networks and Caps Net.The results show that this model has higher accuracy and better robustness in the classification of vehicle types in road surveillance scenarios.2.Aiming at the problem that the existing vehicle classification methods cannot make full use of image information,resulting in low recognition accuracy,a Caps Net infused gradient convolutional features(HOG-C Caps Net)was proposed.HOG-C Caps Net is the model that an extraction method for infused gradient convolutional features is added to the Caps Net.The gradient data in images were calculated by the gradient statistical feature extraction layer,and then the Histogram of Oriented Gradient(HOG)feature map was plotted with the gradient data.The color information of images was extracted by the convolutional layer,and then the HOG-C feature map was plotted with the extracted color information and HOG feature map.After inputting the HOG feature map into the convolutional layer,its abstract features were obtained.The abstract features were encapsulated through a Caps Net into capsules with a three-dimensional spatial feature,which could be calculated by dynamic routing to achieve vehicle classification.Experiments were conducted on the BIT-Vehicle dataset,and the comprehensive evaluation index of HOG-C Caps Net identification reached 98.20%,and the accuracy reached 98.17%,which is superior to the identification of other network structures such as traditional convolutional neural networks and Caps Net.The experiment results show that a better efficiency on vehicle classification of vehicle surveillance can be achieved with the HOG-C Caps Net model.Aiming at the problem of low vehicle recognition accuracy in the context of traffic surveillance,pre-convolutional Caps Net and Caps Net infused gradient convolution features were proposed.Experiments on public datasets have obtained higher accuracy,indicating that the model improve the accuracy of vehicle classification under traffic surveillance.