Algeria licence plate recognition system using faster-RCNN and YOLO models

Abstract

In some institutions, of?ce buildings, or government facilities the ?ow of incoming and outgoing traf?c of people and cars needs to be monitored and recorded for security purposes as well as practicality and automation of entry pass for vehicles. Over the last years, many techniques have been proposed in an attempt to solve the Automatic License Plate Recognition System (ALPRS) problem. These techniques rely mainly on hand-crafted approaches and basic computer vision algorithms such as edge detection with Sobel ?lter. These approaches are not accurate enough for real-world applications, nor are they robust enough to changes in size, shape, and rotation of the license plates. Recently, deep learning techniques have been shown to be a strong tool for solving computer vision and object detection problems, such as ALPRS. In this project, we propose a solution based on convolutional neural networks (CNN). A data set containing 1000 car images has been collected, labeled, and then split into a training set and testing set. The size of this data set would allow for a transfer learning approach and ?ne-tuning of models. In the next step, various models belonging to the “You Only Look Once” (YOLO) CNN and “Faster Region-based CNN” (Faster RCNN) families are trained to perform plate detection task only. Once the models are trained and optimized, they are used to crop images of plates from the original car images. These cropped images are used to train models to perform the digit recognition task, similar to those trained for plate detection. The training process was repeated for different structures and parameters of the models to obtain the best performance possible. Evaluating these models relies on the use of the mean average precision (mAP) used in the original papers of YOLO and Faster-RCNN. The evaluation of the ?nal model (plate detection and digit recognition) relies on the accuracy of performing the identi?cation of the license plate numbers. The end result is an application that achieved an accuracy of 81.36% with real-time video processing capabilities and robust to changes in size, shape, color, and rotation of the license plates. This project provides users of the application with a reliable and practical security tool. It would also supply Algerian academics and software developers with a benchmark data set for further research on the topic and evaluation of future models.