A Study On Identification System Of License Plate Authenticity

Currently,Cracking down on fake license plate is one of the important tasks of traffic system in China.However,identifying the authenticity of license plates is still based on law enforcement officers' visual recognition.In this way,it consumes a lot of manpower and material resources,and the rate of missing recognition is very highIn this paper,we propose an algorithm software system of identifying license plates' authenticity based on image from the perspective of the difference between true and false license plate in fonts.The system which employs image processing and deep learning methods can automatically identify the authenticity of license plateThe main work in the identification system is as follows1.We realize a license plate recognition model based on the non-segmented arbitrary length text recognition network CRNN,which can be carried out on the rough positioning image of the license plate.This model solves the problem that the current license plate number length is not fixed2.We use K-Means-based image binarization and an adaptive connected domain anal-ysis algorithm to accurately locate the license plate.Then,the required key char-acters are extracted by the connected domain analysis method3.In order to position the control points in key character images and find the point correspondence between the reference image and the distorted character image,we propose a key point location model based on convolutional network4.With the RANSAC-based least squares algorithm,the projective transformation be-tween the distorted image and the reference image is determined,and the geometric correction of the key character image is realized.Through the morphological trans-formation,we realize the comparison with the reference image,thereby realizing the true and false discriminating of the license plateExperiments show that the license plate recognition network achieved high accu-racy in multiple test sets of CCPD,and the license plate true and false discriminant model achieved a recall rate of 100%to fake license plates and an accuracy of 91.67%to all samples on a dataset of real license plates.