Abstract:
Automatic License Plate Recognition (ALPR) system is widely used all over the world for traffic vehicle detections. But Bangladesh as one of the populous countries of the world having increasing vehicle demands day-by-day, there is no well-structured ALPR system developed for the detection of vehicles where all the license plates are written in Bangla. Especially, Bangladesh is such a country where congestion is a perennial problem in its capital city Dhaka bearing a huge wastage of money causing incremental traffic delay. Similar problem is observed in all aspects like toll collection, implementation of traffic regulation etc.
Therefore, an effort has been made to develop an ALPR model using MATLAB which could be used to recognize characters from the license plate images through image processing (IP) and Adaptive Neuro-Fuzzy Inference System (ANFIS). As a founding stone of the ALPR system in Bangladesh, 20 frequently used vital characters in the Bangladeshi license plates have been acknowledged to develop the model. Additionally, during the development of the model, some basic assumptions have been made like connected words in Bangla have been treated as distinct characters. Geometric features for individual character have been distinguished using different variable input values which have been used to train ANFIS and later on detecting those characters from the license plate images. There are two basic operations of the ALPR model, first one to detect the license plate region and the second one is to recognize characters through extracting them from that region. ALPR is performed from the stored images in the directory of the program during this research which has been recommended to be converted into real-time captured image detection model for the future studies. For detection of the license plate region, shape, area and background intensity of the license plate have been incorporated to the model. Validation has been done for this model using confusion matrices.
It was observed that for both license plate status and character recognitions, more than 90% accuracy has been attained. Basic operations of the model have also been found satisfactory while executing the program in MATLAB. Relatively greater errors have been identified for the unknown characters as some of those have been recognized as the known characters trained by ANFIS. It can be expected that if all characters are trained through ANFIS, this problem can be mitigated up to a great extent. Also, some complexities have been found during the detection of the license plate region for white colored vehicles as it matches with the background intensity condition of the license plates. Albeit these minor issues, all other operations have been found prominently viable. Therefore, the model can be expected to be a foundation of Bangladeshi ALPR system.
