Real Time Bangla Number Plate Recognition using Computer Vision and Convolutional Neural Network

Abstract

Automatic number plate identification in today's world plays a vital role in vehicle tracking and organization. Our proposed model of automation in the detection and recognizing vehicles through the use of number plate computerization is expected to create a new scope of evolution for large cities. The system can be used for the parking system of motor vehicles, as well as to collect tolls. The detection of the Bangla number plates from different cities and multi-class vehicles is the first step of the proposed system. The number plate detection has been performed with the computer vision approach, and the You Only Look Once v3 (YOLOv3) algorithm. Next, the Tesseract optical character recognition system, in conjunction with the Bangla character recognition model, has been used for vehicle indexing and convolutional neural network for the character recognition from the detected number plate. Numerical results demonstrate that the accuracies of license plate detection for the computer vision and YOLOv3 are 91% and 95%, respectively. For the character recognition, the accuracy for Tesseract and convolutional neural network are 90% if the license plate is detected and cropped successfully and 91.38%, respectively. Finally, our system has been tested using the convolutional neural network method in an environment of real-world where our system's Pi Camera captured video as input, which has a total of 18 different cars. From 18 cars, it has successfully detected 17 cars, which makes our overall system accuracy 88.89%.