Please use this identifier to cite or link to this item: http://nuir.lib.nu.ac.th/dspace/handle/123456789/1468
Title: VEHICLE NUMBER PLATE DETECTION AND RECOGNITION SYSTEM IN BHUTAN
-
Authors: YONTEN JAMTSHO
Yonten Jamtsho
Panomkhawn Riyamongkol
พนมขวัญ ริยะมงคล
Naresuan University. Faculty of Engineering
Keywords: YOLO
Centroid Difference
xgboost classifier
Darknet
Bhutanese license plates
Flip method
Issue Date: 2019
Publisher: Naresuan University
Abstract: Bhutan is a small landlocked country sandwiched between two giants of the world; China in the north and India in the south, east and west. Bhutan had remained isolated from the outside world since 1970s when it was opened to foreigners. Bhutan is one of the last countries to introduce television and internet. With the rapid development and urbanization, more people have begun to migrate from rural to urban areas due to more opportunities and availability of facilities. As the number of people increases in the urban areas, there has been a problem of job opportunities, traffic congestion and human health problem. Apart from this, with the increasing number of vehicles in the country, it has become crucial to automate the traditional way of managing traffic since it creates more traffic congestion and needs more human resources to manage the traffics. The study aimed to develop an automatic license plate recognition system for a Bhutanese license plate and an algorithm for the extraction of features from the characters.  In the study, the detection of vehicle and localization of license plate was based on the latest state-of-the-art YOLO (You Only Look Once) object detector. An automatic license plate localization from the vehicle was put forward to reduce the number of false positives generated by the signboard and other objects since they look similar to the license plate. Once the license plate was extracted, several image preprocessing steps such as automatic cropping of Bhutanese scripts, noise removal and contrast enhancement was applied to improve the quality of the image before segmenting the characters. Also, a white pixel counting with inversion method was introduced to handle the taxi (BT) license plate since it gives different output after the thresholding operation. For character segmentation, a connected component analysis (CCA) was proposed, which labels all the unique components in the binary image. After segmenting the characters, three statistical concepts: standard deviation, mean absolute deviation and modified z-score were used to remove the unwanted characters segmented during the segmentation phase. After that, the character classifier was modelled by extracting the 4948 features from 17 characters. For feature extractions, a centroid difference(CD) and flip methods along with Hu's moments were proposed to handle those characters having similar Hu's moments. The extracted features were trained and tested using XGBoost classifier.  In the license plate localization, 1050 vehicle images were given to the proposed method. From the experiments, the overall vehicle detection accuracy was 98.5%, and the license plate localization accuracy was 97.9%. The extracted license plates from the localization step were evaluated to check the performance of the segmentation method and achieved a segmentation accuracy of 96.1%. All the correctly segmented license plates were tested to monitor the performance of the recognition method and resulted in overall recognition accuracy of 92.9%. This research is first of its kind to use Bhutanese datasets for the development of ANPR technology.  The proposed method achieved high accuracy and outperformed some methods discussed in the literature. 
-
Description: Master of Engineering (M.Eng.)
วิศวกรรมศาสตรมหาบัณฑิต (วศ.ม.)
URI: http://nuir.lib.nu.ac.th/dspace/handle/123456789/1468
Appears in Collections:กลุ่มวิทยาศาสตร์และเทคโนโลยี

Files in This Item:
File Description SizeFormat 
61062946.pdf5.6 MBAdobe PDFView/Open


Items in NU Digital Repository are protected by copyright, with all rights reserved, unless otherwise indicated.