Design And Implementation Vehicle Multi-label Recognition Algorithm Based On SSD

The year 2020 is a prosperous and strong time for China to enter into a well-off living standard in an all-round way.At this time,the automobile has become a daily tool for Chinese people.The ever-increasing number of vehicles has brought tremendous pressure to traffic.In order to alleviate traffic pressure,the collection of vehicle information is very important for traffic managers.With the rapid development of machine learning and artificial intelligence,obtaining vehicle target information through image processing of deep learning has solved this problem well.With the rapid development of machine learning and artificial intelligence,the problem can be solved through the image processing of deep learning to obtain the information of vehicle target.The vehicle object is composed of a variety of attributes,such as color,model,license plate and other attributes can be used as the basis for vehicle object identification.Each attribute of the vehicle object is used as a type of label.Generally speaking,identifying each vehicle attribute requires multiple deep learning network algorithms,which greatly increases the calculation cost.Therefore,the research of multi-label recognition of vehicles is very meaningful.Due to the rapid development of target detection and recognition network,a series of target detection and recognition networks derived from Convolutional Neural Network(CNN)have been applied to road traffic vehicle detection and recognition and achieved very good results.SSD(Single Shot multibox Detector)algorithm meets the requirements of vehicle detection in accuracy,real-time and multi-scale detection.This paper focuses on vehicle detection and multi label recognition based on all the advantages of SSD,and completes a vehicle multi-attribute detection system according to the traffic demand.The specific work of this paper is as follows:(1)In this paper,in order to meet the particularity of SSD algorithm for vehicle multi-label recognition,it is necessary to make a training data set which is different from the conventional voc2007 data set.The conventional voc2007 data set for target detection is for single image and multiple targets,and each target has only one label.The training data set used in this paper is to label each vehicle target in a single image with multiple labels.In this paper,two attributes of vehicle(vehicle type,body color)are used as multi-label of vehicle.Since the data set labeled by hand is only 1000,data enhancement is needed to improve the accuracy and robustness of the algorithm.(2)Changing the convolution layer of SSD network to obtain category of confidence.Through different 3 * 3 convolution kernels,there are two kinds of categories of confidence and one position regression for each prediction frame.The loss function of it is also improved,and a weighted sum of two kinds of categories of confidence loss function and a position regression loss function is made to obtain the loss function of the whole network,Introduce the coefficient ? of the color category loss function,and optimize the ? parameter through experiments to obtain the optimal ?= 2.(3)In order to meet the needs of road traffic vehicle attribute detection,this paper deploys multi-label detection algorithm based on SSD to vehicle detection system.Through the vehicle image captured by the front-end camera,the attributes of each vehicle target in the image can be identified and displayed on the WEB page.