A Real-time End-to-End Bangladeshi License Plate Detection and Recognition System for All Situations Including Challenging Environmental Scenarios

Automated License Plate Detection and Recognition (ALPDR) system is a technology that automatically detects and recognizes the license plate from a video or an image. An ALPDR system is usually a 3 steps process; license plate detection, segmentation of characters, and finally recognition. For an effective ALPDR system, every step of the process should be highly accurate and robust. This is a very active research field and many researchers are working towards building an effective ALPDR and proposed various recognition systems for Bangladeshi license plates. In this paper, we propose a real-time end-to-end ALPDR system for Bangladeshi vehicles. Our proposed system has three sub-systems which are Detection, Segmentation, and Recognition. For the detection of license plate, we have used a real-time detection model called You Only Look Once (Yolov4) and achieved a very high accuracy of 99.89%. For segmentation, we have implemented an existing greedy graph-based segmentation algorithm which has 99.52% overall accuracy. For recognition, we have developed a Convolutional Neural Network (CNN) based model which achieved a 99.33% fl-score and is quite generalized. Moreover, we have created the largest Bangladeshi-Licence-Plate dataset consisting of 5087 diverse images of vehicles captured from the streets of Bangladesh.