Research On Vehicle Color Recognition Based On Deep Learning

With the development of intelligent transportation,there are more and more vehicles in the urban road.With the occurrence of many traffic accidents,in order to better monitor and track vehicle information,it is necessary to carry out vehicle identification.A lot of research work has been accumulated for license plate recognition,but there are some problems in license plate recognition,such as license plate damage,etc.as a reliable and significant vehicle feature,vehicle color can also provide useful information for vehicle recognition,monitoring,tracking and other aspects,which can solve the problem of obtaining vehicle information only by license plate recognition in complex scenes.Vehicle color recognition is a technology to determine the vehicle color category in the image or video according to the input vehicle image or video,which has been widely used in the fields of public security law enforcement,vehicle tracking,etc.However,there are many kinds of vehicle colors in the real scene,and the color is easy to be interfered by the external environment,resulting in color error,which has a certain impact on the accuracy of vehicle color recognition.Aiming at the problems existing in the task of vehicle color recognition,this paper proposes a complete process of vehicle color recognition based on the traditional deep learning network to improve the accuracy of vehicle color recognition.In this paper,firstly,according to the image captured by the bayonet monitoring,the color data set of the self built vehicle is built,and the vehicle body and face images in other scenes are collected as the two test sets of this paper to test the model.Then,by comparing the recognition performance of several commonly used deep learning networks on the vehicle color data set built in this paper,the basic network is improved with the idea of residual network,and attention mechanism and multi-scale convolution operation are introduced into the network respectively to realize the good fusion of the following features on the network,so that the network can learn more abundant deep semantic information and shallow layer Edge texture information,and update the loss function in the original network to improve network performance.Because the vehicle color image is collected by the bayonet camera,it is very easy to be affected by the light,fog and other external environment.At the same time,if the performance of the data acquisition equipment is not good,the collected image will also have the problem of low resolution.After the data is trained by the model,the network classifier will output the color category,and the model has the situation of false classification.In view of the above problems In this paper,we first preprocess the vehicle color image in the data set,and then propose a method of converting RGB image to HSV color space,counting the value range of ten kinds of colors in HSV space,binary processing vehicle image,counting the area occupied by pixels in the image,and distinguishing color types according to the area threshold.Through the experimental comparison and analysis of the results of the proposed vehicle color recognition process,we can see that the recognition accuracy of the proposed network,which combines the residual thought,attention mechanism and multi-scale feature extraction,is increased by about 1.5% compared with the residual network,and the memory occupation of the network in this paper is reduced by 20 m compared with the residual network,which is conducive to the application of the model in practice 。 The image processing of misclassification proposed later can be distinguished by the area threshold with large difference in order of magnitude,which provides a reasonable solution for the model to recognize the wrong image.