Abstract

License plate detection and character recognition pose challenges due to environmental
sensitivity, such as lighting, dust, and the impact of the chosen font type on recognition
tasks. Automatic License Plate Detection and Recognition (ALPR) are crucial
in practical applications such as traffic control and parking, vehicle tracking, toll
collection, and law enforcement. While much research has been done using image
processing and machine learning algorithms, deep learning methods need further
exploration due to their recent advances in reliable performance in various scenarios.
Moreover, current proposals are limited to specific regions and dataset applicability.
This study has a dual focus: firstly, we suggest utilizing a Deep Learning technique,
specifically using Faster R-CNN for the license plate detection task and a CNN-RNN
model with CTC loss, and a MobileNet V3 backbone for recognition task. We also
utilized You Only Look Once (YOLO) for license plate detection and recognition tasks.
Secondly, we aim to assess font features within the LP context. This work uses Brazilian
dataset and datasets from two different provinces in Canada and two different states in
the United States of America, including Ontario, Quebec, California, and New York
State. We suggest employing an adaptive algorithm based on Faster R-CNN and CTC
network along with YOLO, fine-tuned with optimized parameters to improve its
effectiveness using two different approaches, including domain generalization.
Alongside presenting the recall ratio findings, this study will perform a thorough error
analysis to gain insights into the nature of false positives. The proposed model
demonstrated a commendable recall ratio of 94% using a single YOLO network.
Specific fonts pose readability challenges for humans, while others present difficulties
for computer systems regarding recognition. In this study, we provide five sets of
outcomes for font assessment: results about font anatomy and those related to the
recognition of commercial products. The font anatomy analysis focuses on five specific
fonts: Driver Gothic, Dreadnought, California Clarendon, Zurich Extra Condensed, and
Mandatory. Additionally, we assess the impact of these fonts in the context of a dataset
made of five different license plates using a commercial product, OpenALPR. The font
anatomy findings unveil significant confusion cases and quality features associated with
chosen fonts.