MALAYSIAN AUTOMATIC LICENSE PLATE RECOGNITION USING SINGLE-SHOT OBJECT DETECTION MODEL AT LOW VISIBILITY AND UNCONSTRAINED ENVIRONMENT

Abstract

Automatic license plate recognition in the outdoor environment is challenging due to many factors that affect their performance. These include unconstrained environment conditions and variations in the license plate (LP) designs. In an open-world situation, the problem becomes more complicated due to non-compliance with the standard specifications and the existence of the special LP. To overcome this problem, we proposed a huge Malaysian license plate dataset and employed a high-accuracy single-shot object detection network YOLOv5 for the plate detection and character recognition tasks in a hierarchical manner. The dataset was collected by extracting video frames from cameras installed at different toll plaza booths day and night under various weather conditions. To capture large variations of license plate styles, camera viewpoints, lighting conditions as well as character font templates, we devised several augmentation strategies specifically tuned to address these challenges. The large dataset of 58,000 images was used to train the YOLOv5 model with 1200 epochs. The system performed very well on the challenging test dataset achieving a 95.96% accuracy rate.