Deep learning based car license plate recognition

Abstract

Currently, there is a lack of license plate recognition systems that are lightweight and fast, while still being sufficiently accurate for practical purposes. In this project, we explored various methods to adapt convolutional neural networks which fulfil the above requirements for usage on Singaporean license plates. In particular, we carried out pre-training and fine-tuning of LFFD, such that it reached an average precision of 98.99% for license plate detection. In addition, we modified the backbone architecture of LPRNet for it to handle single-row and double-row license plates, and tried out various data augmentations to improve its accuracy, such that it obtained an accuracy of 93.79% for license plate recognition. We then combined the two models to create a system that is able recognised the license plate number given an image of a Singaporean vehicle. This system is lightweight, having only a total size of 7.6 MB, and fast, taking 82 ms to process an image on average. It also has a decent recognition accuracy of 86.04%.