Research And Application Of License Plate Recognition Based On Deep Learning

As an important and popular research direction in Intelligent Transportation System(ITS),license plate recognition has been widely used in various traffic situations to alleviate traffic congestion,road traffic monitoring,traffic management,and illegal capture.It has played a positive role and has high social research value.With the explosive growth of car ownership,how to obtain license plate information faster and more accurately has become a hot research topic.This paper takes intelligent transportation as the starting point and stands on the traditional algorithm theory to study the algorithm of vehicle license plate recognition based on deep learning.Starting from the traditional license plate recognition algorithm,this paper studies the characteristics of domestic license plates and common image features,pre-processes the license plate image,removes noise,and improves the accuracy of license plate detection and recognition.In this paper,the HOG(Histogram of Oriented Gradient)local feature descriptor is used to extract the license plate information.The traditional support vector machine(SVM)algorithm is used to complete the license plate recognition,and the two-class SVM is used.License plate location,multiple binary SVMs complete license plate character recognition.Although this method obtains a high accuracy of license plate location,the accuracy of character recognition is only 90%.Aiming at the shortcomings of traditional SVM algorithm in license plate recognition,this paper uses MobileNet,a lightweight network model developed by Google,to build a Convolutional Neural Network of MobileNet-SSD,using the SSD(Single Shot multibox Detector)target detection network.Depth-separable convolution reduces the SSD convolutional network parameter size and multiply-accumulate calculations.The method abandons the cumbersome steps of the traditional license plate recognition,realizes the end-toend license plate recognition,avoids the accumulation of intermediate errors,and completes the segmentation recognition of the license plate characters.After experimental testing,the method has higher recognition accuracy in complex environments,and the accuracy of character comprehensive recognition reaches 95%.Finally,this paper compares the traditional SVM algorithm with the deep learning MobileNet-SSD convolutional network algorithm for comprehensive license plate recognition test.A simple GUI test interface was implemented,and the license plate images of the typical complex environment were tested separately using the license plate image of the Chinese Vehicle Parking Dataset(CCPD).The experimental results show that in the general environment,the recognition accuracy of the two methods is not much different.However,in the complex environment,the experimental results show that the license plate recognition based on MobileNet-SSD has strong robustness,especially for the recognition of long-distance license plates,and has strong adaptability to noise environment.