Synthetic image generation for training deep learning-based automated license plate recognition systems on the Brazilian
- PDF / 3,518,177 Bytes
- 21 Pages / 439.37 x 666.142 pts Page_size
- 29 Downloads / 218 Views
Synthetic image generation for training deep learning-based automated license plate recognition systems on the Brazilian Mercosur standard Gilles Silvano1 · Vinícius Ribeiro1 · Vitor Greati1 · Aguinaldo Bezerra1 · Ivanovitch Silva1 · Patricia Takako Endo2 · Theo Lynn3 Received: 30 April 2020 / Accepted: 7 September 2020 © Springer Science+Business Media, LLC, part of Springer Nature 2020
Abstract License plates are the primary source of vehicle identification data used in a wide range of applications including law enforcement, electronic tolling, and access control amongst others. License plate detection (LPD) is a critical process in automatic license plate recognition (ALPR) that reduces complexity by delimiting the search space for subsequent ALPR stages. It is complicated by unfavourable factors including environmental conditions, occlusion, and license plate variation. As such, it requires training models on substantial volumes of relevant images per use case. In 2018, the new Mercosur standard came in to effect in four South American countries. Access to large volumes of actual Mercosur license plates with sufficient presentation variety is a significant challenge for training supervised models for LPD, thereby adversely impacting the efficacy of ALPR in Mercosur countries. This paper presents a novel license plate embedding methodology for generating large volumes of accurate Mercosur license plate images sufficient for training supervised LPD. We validate this methodology with a deep learning-based ALPR using a convolutional neural network trained exclusively with synthetic data and tested with real parking lot and traffic camera images. Experiment results achieve detection accuracy of 95% and an average running time of 40 ms. Keywords Mercosur license plates · Automated license plate recognition · License plate detection · Number plate detection · Smart cities · Deep learning · Synthetic data
B
Gilles Silvano [email protected] Vinícius Ribeiro [email protected]
1
Department of Informatics and Applied Mathematics, Federal University of Rio Grande do Norte, Natal, Rio Grande do Norte, Brazil
2
University of Pernambuco, Pernambuco, Brazil
3
Irish Institute of Digital Business, Dublin City University, Dublin, Ireland
123
G. Silvano et al.
1 Introduction License plates, also referred to as vehicle registration plates or number plates, are the primary source of vehicle identification data. Since their introduction at the turn of the 19th century, the main feature of a license plate has remained constant i.e. a numeric or alphanumeric identifier that uniquely identifies the vehicle or the vehicle owner within the issuing region’s vehicle register. Initially, this identifier served two functions - (i) fast and accurate vehicle identification, and (ii) compliance with vehicle registration laws [6]. To achieve this, license plates are, typically, fixed directly to a vehicle, or indirectly to a plate frame, and are designed to conform to standards and technical specifications for legibility and re
Data Loading...
