Research On Enlarged License Plate Detection And Recognition Based On Deep Learning

With the continuous improvement of urban construction and the rapid development of society,building a complete intelligent transportation system has gradually become a necessary part of building an intelligent city.The management of large vehicles in intelligent transportation system is as important as that of ordinary cars,but it is more difficult.In terms of obtaining the identity of the vehicle,due to the influence of license plate location,shooting angle and stains,the current license plate detection methods are not good enough at getting license plate information.However,an effective way to solve this problem is to obtain the license plate number by detecting the enlarged license plate painted at the rear of the truck.Because the character spacing of the enlarged license plate is relatively large,the current text detection methods are prone to false detection.For this reason,this thesis focuses on how to detect and identify the enlarged license plate with high accuracy by investigating the separate two-stage detection and recognition method and the end-to-end detection and recognition method respectively.The main work and contributions of this thesis could be summarized as follow:1.This thesis improves text detection method based on the character region for the enlarged license plate detection scenario,and implements a recognition module based on character detection.On the basis of character area score map and character affinity score map,a character similarity score map based on character height and central point vertical axis height is designed for the enlarged license plate.The weighted sum of character similarity score map and character affinity score map is used to comprehensively evaluate the relationship between adjacent characters,so as to alleviate the problem of false detection caused by large character spacing.As can be seen from the experimental results,compared with the original method,the detection accuracy of this method for enlarged license plate is improved by 3.74%,which proves the effectiveness of this method.2.In this thesis,an end-to-end method for detection and recognition of enlarged license plate is proposed.The method introduces an attention module consisting of spatial attention and channel attention to strengthen the contextual connection of features.A multi-branch network of overall detection,mask detection and character detection recognition is constructed to detect the overall and characters of the enlarged license plate.At the same time,an evaluation module is designed according to the character characteristics and naming rules of the enlarged license plate,and the evaluation score of the evaluation module is jointly evaluated with the confidence score of the overall detection.The experimental results show that for the detection of enlarged license plate,the accuracy of this method is 96.34%,which verifies the feasibility of this method.