Paper Info
Title
Brazilian Mercosur License Plate Detection: a Deep Learning Approach Relying on Synthetic Imagery
Abstract
Automated license plate recognition (ALPR) technology is a powerful technology enabling more efficient and effective law enforcement, security, payment collection, and research. A common license plate standard was adopted by the member states of the Mercosur trading bloc (Argentina, Brazil, Paraguay and Uruguay) and consequently requires an upgrade to the ALPR software used by law enforcement and industry. Due to the scarcity of real license plate images, training state-of-the-art supervised detectors is unfeasible unless data augmentation techniques and synthetic training data are used. This paper presents an accurate and efficient automated Mercosur license plate detector using a Convolutional Neural Network (CNN) trained exclusively with synthetic imagery. In order to obtain the synthetic training data, Mercosur license plates were faithfully reproduced. Digital image processing techniques were employed to reduce the domain gap and a CNN with basic image manipulation was used to embed the artificial licensed plates in to realistic contexts. The trained model was then validated on real images captured from a parking lot and a publicly available traffic monitoring video stream. The results of experiments suggest detection accuracy of about 95% and an average running time of 40 milliseconds.
Year
DOI
Venue
2019
10.1109/SBESC49506.2019.9046091
2019 IX Brazilian Symposium on Computing Systems Engineering (SBESC)
Keywords
DocType
ISSN
mercosur license plates,automated license plate recognition,license plate detection,smart cities,deep learning,synthetic data
Conference
2324-7886
ISBN
Citations 
PageRank 
978-1-7281-6319-2
0
0.34
References 
Authors
17
7
Name
Order
Citations
PageRank
Vinícius Ribeiro100.68
Vitor Rodrigues Greati200.34
Aguinaldo Bezerra300.68
Gilles Silvano400.68
Ivanovitch M. D. Silva58713.56
Patricia Endo612125.32
Theo Lynn711622.40