Number plate detection using deep learning techniques

Abstract

Automatic number plate recognition (ANPR) is a sub-area of machine vision that has been investigated extensively due to its practical uses and benefits. ANPR has been implemented by the police in several countries and is used to detect, deter, and disrupt criminality. These systems are used in conjunction with databases containing car registrations pertaining to stolen vehicles and vehicles involved in crimes. Furthermore, ANPR can be implemented to enforce automatic vehicle access control. This allows enforcement of low emission vehicle zones in city centres to combat worsening air quality. This study aims to develop an effective number plate recognition solution for Maltese number plates. The solution must be adaptable to many varying scenarios in terms of distance from the vehicle, angle, lighting conditions, number plate type, and other variables. A number plate detection model using YOLOv4 and a custom character recognition module using Tesseract optical character recognition is used to achieve this. YOLOv4 is a single-shot object detection method that uses the Darknet framework for its training and number plate detection. In this project, three separate models were trained using YOLOv4, and their accuracy was evaluated on a Maltese number plate test dataset. Two of the models were trained on individual data sets collected. One dataset is specific to Belgian number plates, and the other is a more varied data set collected from Google OID v6. The third model was trained using the combination of both data sets. The differences in the efficacy of these three models were observed through their detection accuracy on the Maltese test data set. The accuracy of the models is based on the mean average precision (MAP), and the intersection over union (IoU) obtained. For the recognition segment of this work, a suitable project was selected, which uses Tesseract OCR to perform the recognition task. Tesseract OCR performs best when the characters are preprocessed to make them more easily readable to Tesseract. Therefore, the chosen project implements several preprocessing steps to the number plate detection. However, the accuracy of the OCR was observed to be suboptimal on the Maltese number plate test set. Hence two new methods were refined to optimise and augment the preprocessing steps to improve the accuracy of the final recognition output. The character recognition modules’ accuracy was compared based on the Jaro Similarity and Levenshtein distance of their outputs on the Maltese test dataset.