YOLO-Based Three-Stage Network for Bangla License Plate Recognition in Dhaka Metropolitan City
| Autor: | Mahedi Hasan, Sohaib Abdullah, Sheikh Muhammad Saiful Islam |
|---|---|
| Rok vydání: | 2018 |
| Předmět: |
Computer science
Intersection (set theory) business.industry Character (computing) Deep learning ComputingMethodologies_IMAGEPROCESSINGANDCOMPUTERVISION Viewing angle Convolutional neural network language.human_language Bengali language Computer vision Artificial intelligence Unavailability business License |
| Zdroj: | 2018 International Conference on Bangla Speech and Language Processing (ICBSLP). |
| DOI: | 10.1109/icbslp.2018.8554668 |
| Popis: | Real-time Automatic License Plate Recognition (ALPR) is a challenging task in computer vision since license plates in the real-world scenarios are difficult to recognize due to the presence of various factors like transparent background, occlusion, existence of multiple plates in an image, variance in illumination and viewing angle, etc. In most of the previous works related to Bangla license plate, images used for ALPR were captured in ideal conditions due to unavailability of dataset representing non-ideal conditions. In this paper, we present an approach for real-time Bangla license plate detection that is robust to enormous variations in complex real-world environment. A novel license plate recognition method is also presented making it a complete end-to-end deep learning-based ALPR system. We introduce a dataset by collecting 1, 500 different Bangladeshi vehicular license plate images that are captured manually from the street resembling various real-world scenarios. In this work, we have employed YOLOv3 algorithm to successfully localize the license plate and recognize the digits. To recognize the character, we have also built a Bangla scene character dataset containing more than 6, 400 characters, with which we have trained a ResNet-20-based deep Convolutional Neural Network (CNN). Our proposed method achieves more than 85% Intersection over Union (IoU) in digit recognition. The ResNet-20-based CNN model achieves 92.7% accuracy in recognizing the Bangla character present in the license plate. |
| Databáze: | OpenAIRE |
| Externí odkaz: |
|