• deep learning
• automatic plate recognition
• object detection
• faster r-cnn
• yolo

In the class of the problems that are part of Computer Vision and, in particular, in the Object Detection domain, Automatic Plate Recognition is a common field of research at the moment, probably also because of the increasing real-world applications that could benefit from it.
In the last years and with the growing development of new Deep Learning technologies, lots of steps have been made. However, the challenges to face are many, specifically considering the multiple factors that can affect the performances of these systems, so there is still a lot of research to be done.
Considering this context, the objective of this thesis work is to develop a system that is able to detect license plates from low quality pictures in all the possible situations. The final system will be integrated in a bigger architecture that includes the database of a famous car rental company, in order to guarantee the privacy of all the sensible data that regards the available cars.

After a careful analysis of the existing approaches, a lack of available data to train state-of-the-art Deep Learning models has been highlighted. As a result, a dataset was created by collecting and annotating different images of cars and their license plates from european countries to finetune some existing architectures. Subsequently, some experiments on the given test set from a sample of the database of the car rental company have been performed. In order to check the detection capability of the final system, an additional test on a well known dataset in the literature has been carried out to compare the results with the existing solutions.
Eventually, after analyzing the outcomes, possible improvement steps are proposed.