Iranian License Plate Recognition Using a Reliable Deep Learning Approach

Abstract

The issue of Automatic License Plate Recognition (ALPR) has been a challenging one in recent years because of weather conditions, camera angle, lighting and different license plate characters. Due to advances in deep neural networks, it is now possible to recognize Iranian license plates using specific neural networks. The proposed method recognizes license plates in two steps. First, license plates are detected through the YOLOv4-tiny model, which is based on Convolutional Neural Network (CNN). Secondly, Convolutional Recurrent Neural Network (CRNN) and Connectionist Temporal Classification (CTC) are applied to recognize the license plate characters. For labels, one string of numbers and letters is enough without segmenting and labeling each separately. The proposed method boasts an average response time of 0.0074 seconds per image and 141 frames per second (fps) in the Darknet framework and 0.128 seconds per image in the TensorFlow framework for the License Plate Detection (LPD) part. This method has been proven to provide a highly accurate model with minimal storage space requirements, using less than 2MB for the Character Recognition (CR) model. There was an average accuracy of 75.14% and a response time of 0.435 seconds for the end-to-end process. The released code is available through GitHub.