Abstract

The current advancements in machine intelligence have expedited the process of recognizing vehicles and other objects on the roads. The License Plate Recognition system (LPR) is an open challenge for many researchers to develop a reliable and accurate system for automatic license plate recognition. Several methods including Deep Learning techniques have been proposed recently for LPR, yet those methods are limited to specific regions or privately collected datasets.
In this thesis, we propose an end-to-end Deep Convolutional Neural Network system for license plate recognition that is not limited to a specific region or country. We apply a modified version of YOLO v2 to first recognize the vehicle and then localize the license plate. Moreover, through the convolutional procedures, we improve an Optical Character Recognition network (OCR-Net) to recognize the license plate numbers and letters.
Our method performs well for different vehicle types such as sedans, SUVs, buses, motorbikes, and trucks. The system works reliably on images of the front and rear views of the vehicle, and it also overcomes tilted or distorted license plate images and performs adequately under various illumination conditions, and noisy backgrounds. Several experiments have been carried out on various types of images from privately collected and publicly available datasets including OPEN-ALPR (BR, EU, US) which consists of 115 Brazilian, 108 European, and 222 North American images, CENPARMI includes 440 from Chinese, US, and different provinces of Canada and UFPR-ALPR includes 4500 Brazilian license plate images; images of those datasets have several challenges: i.e. single to multiple vehicles in an image, license plates of different countries, vehicles at different distances, and images taken by several types of cameras including cellphone cameras. Our experimental results show that the proposed system achieves 98.04% accuracy on average for OPEN-ALPR dataset, 88.5% for the more challenging CENPARMI dataset and 97.42% for UFPR-ALPR dataset respectively, outperforming the state-of-the-art commercial and academics.