A Robust License Plate Detection and Recognition Framework for Arabic Plates with Severe Tilt Angles

Abstract

This paper addresses the challenge of accurately detecting and recognizing Arabic license plates, particularly those subjected to severe tilt angles. It presents a robust license plate detection and recognition framework that consists three main steps: plate detection and segmentation, plate perspective correction, and vehicle number recognition. In the first step, a mask R-CNN model is used to detect the plate location, providing pixel-wise labels of identified plates' areas. Following this, a perspective correction technique is used to obtain a clear and rectangular image of each license plate in the image. Lastly, the framework employs a Bidirectional Long Short-Term Memory (Bi-LSTM) model for accurate vehicle number recognition. The framework's efficacy is demonstrated through its application to build a plate recognition system tailored for Egyptian license plates. The system was tested on a dataset collected from campus gate cameras at Zewail city of science and technology, achieving a character accuracy of 97%.