A Real-Time Automatic Kurdistan Numberplate Recognition System

M. Ashir, Abubakar (2022) A Real-Time Automatic Kurdistan Numberplate Recognition System. Eurasian Journal of Science and Engineering, 8 (1).

[img] Text (Research Article)
eajse.v8i1p126.pdf - Published Version

Download (1MB)
Official URL: https://eajse.tiu.edu.iq/index.php/volume-8-issue-...

Abstract

The current paper presents a real-time implementation of a numberplate recognition system for Kurdistan region of Iraq. Automatic numberplate recognition systems (ANPR) have played a key role in many places in enforcing regulations on safe driving and preventing vehicular theft, identity establishment and many other applications. Though there are number of such systems available, their accuracy and speed of execution at inference time is still a challenging issue. To detect a numberplate from a fast-moving vehicle or from a high-resolution camera input, an extremely fast algorithm is required capable of processing tenth of frames per second. This is a huge challenge for many systems and often a compromise is made between the accuracy of detection and execution speed. This work proposes and implement Automatic numberplate recognition system with high accuracy and capable of processing over 50 frames per second at image resolution of 1920×1080 on a raspberry pi B 4 processor. The proposed approach has two major parts: numberplate region localization and character recognition or extraction from the localized region. We used standard machine-learning approach to detect the region of interest using Haar-like features algorithm as feature extractor and Adaptive Boosting (AdaBoost) algorithm to train a cascade of weak learner’s classifiers for classification. After detection of the numberplate region from the input image, an optical character recognition algorithm (Tesseract) is used to extract the characters from the image for display and other use. Tesseract is a machine-learning based OCR algorithm which was pretrained with many languages and made available by google. To increase the detection accuracy, we proposed a masked training approach. The masked training approach uses masked positive samples as negative samples during the training. We also investigated the effect of using different boosting optimization techniques on the overall accuracy of the system. The overall accuracy and inference speed has greatly been improved when tested on a raspberry pi 4 B hardware.

Item Type: Article
Uncontrolled Keywords: Numberplate, Haar-like Features, Cascaded Classification, AdaBoost, OCR
Subjects: Engineering > Computer engineering
Divisions: Eurasian Journal of Science and Engineering > VOL 8, NO 1 (2022)
Depositing User: TIU ePrints Admin
Date Deposited: 24 Oct 2022 13:08
Last Modified: 05 Dec 2022 08:02
URI: http://eprints.tiu.edu.iq/id/eprint/1080

Actions (login required)

View Item View Item