Automatic Multi-Style Egyptian License Plate Detection and Classification Using Deep Learning

Abstract

Detecting Plates in road traffic streams is the most important phase in Automatic License Plate Detection and Recognition (ALPR). ALPR is used in several applications, like motorway road tolling, border control, and parking systems. Recently several techniques were proposed to solve this problem. High prediction accuracy and fast ALPR is necessary. Egyptian license plates (ELP) are considered one of ALPR's difficult types due to the variety of plate types, ranging from standard to non-standard, including Arabic or Latin fonts. In this research, the Deep Learning (DL) technique is proposed to detect and classify Egyptian license plates. The proposed system can handle both main Standard and Nonstandard plates simultaneously with high accuracy and acceptable speed. The proposed system overcomes the main challenges; the large variance of ELP styles and the lack of a reasonable size of data samples. The algorithm is divided into two main stages. The first stage consists of plate detection using deep learning detection methods based on state of art Yolov3. Then the second stage consists of applying deep learning classification methods to classify the detected plate into one of the license plates styles. The proposed algorithm achieved 99.2% detection accuracy for standard plate style and 96.8% for non-standard plates which is a 5% improvement over the state-of-the-art techniques. For the classification task, we used two techniques: Siamese network and multi-task learning. The best result was obtained using the Siamese network achieving classification accuracy of 98.9% for standard and 97.26 <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">%</sup> for non-standard at 33 fps using a low range GPU.