Object Localization and Detection for Real-Time Automatic License Plate Detection (ALPR) System Using RetinaNet Algorithm

Abstract

Automatic License Plate Detection (ALPR) has become important for various fields such as access control and traffic monitoring. The most important steps in ALPR are the license plate localization and detection. Factors affecting the accuracy of the localization and detection of the license plate are due to the uncontrolled environment factors such as lighting conditions, blurred images, occlusion, non standardized plates, irregular image acquisition. In recent years, many researches focus on the use of deep learning network to obtain a more accurate localization and detection of license plate for a better accuracy and effective ALPR system. The main reason for using these types of networks is due to the capability of the networks to automatically localize and detect the license plates in uncontrolled environment. This paper presents the use of one of the deep learning technique which is RetinaNet algorithm and the Convolution Neural Network (CNN) ResNet-50 to localize and detect the license plates taken from video images. The videos were taken in real time both during the daytime and night-time. Over 3,000 images are extracted over 7 days of video recording which are trained using the network. The results show the accuracy of 97.82% which is an improvement compared to a Vertical Projection and Cascade Classifier implemented on this application.