Deep learning based license plate recognition system for controlled transitions

Automatic Lisence Plate Recognition is a widely needed system in many places such as highways, car parks, terminals, public institutions and organizations. In this study, plate recognition system is implemented. Plate recognition systems consist of three main stages. These stages are plate detection, character segmetation and character recognition. The Tensorflow Object Detection API is used for the plate and character detection process in the study. The multilayer convolution neural networks are used for character recognition. These three main operations take place in the following order: First, the plate detection process is performed on the input image of the system. Then, detected plate is subjected to re-sizing and filtering processes and made ready for character recognition. In the second stage, the characters on the plate are detected. In the third and last stage, the recognition process of the detected characters is performed. The system runtime takes on average 5.25 seconds on a computer with the i5-8250 CPU 1.80 GHz processor. When the three main stages of system performance are analyzed separately: plate detection process performance is 99.06%, character detection performance is 95.03% and character recognition performance is 90.03%. But this modular performances can not been obtained for overall system performance. The main reasons of this performance decrease are plate slope and environmental factors. In the future works, it is aimed to incerase the performance of the system by overcoming this problems.