License Plates Classification Based on Deep Learning Technique

Resumen

Multiple agencies worldwide have quickly and widely adopted the recognition of a vehicle
license plate technology to expand their ability in investigative and security matters. License
plate helps detect the vehicle's information automatically rather than a long time consuming
manually gathering for the information. In this thesis, a car plate recognition proposed based
on the deep-learning technique. It includes two phases: localization, classification. In the first
phase, the extraction of the vehicle plate is performed through four steps: preprocessing, plate
detection, connected component analysis and Hough transform. In the second phase, the use
of deep learning via Convolutional Neural Networks (CNN) is established as multiclass
model. The proposed system is tested, and the achieved results are as follows. In the
localization phase, about 600 images are tested for three countries (Belarus, Armenia and
Hungary) under different weather conditions. The corrected number of the extracted plates is
591, with an accuracy of 99.12 % and 100% for the deskewed plate. In the classification
phase, the multi- class classification of the license plate is performed for the three countries
with a dataset of 560 images. Each class has 200 images that are divided into 140 for training
and 60 for validation. The optimal settings for a classification model are found with the use
of a categorical cross-entropy and Adam's optimizer with a learning rate of 0.0001 is used.
For both the training and validation test sets, the model achieved an accuracy of 99.80% and
then 100% respectively. After testing on 80 images, the classification accuracy was found to
be 95% in overall. The duration of the training is 15 minutes.