Robust Automatic License Plate Recognition Using Synthetic Data and Transformer-Based Deep Learning Model

Car license plates in Oman feature a unique vertical arrangement of Arabic and English letters. This distinctive format, combined with the often low-resolution images, presents challenging condition for existing Automatic License Plate Recognition (ALPR) systems which mostly assume horizontally oriented plates with monolingual characters. This frequently leads to issues such as missed image regions or inaccurate segmentation, resulting in incorrect alphanumeric sequences. Traditional deep learning models used in ALPR systems typically need extensive hand-labeled real-world datasets to work effectively under these conditions. To overcome this, our study introduces a robust method that uses 86,000 synthetically generated images. The proposed method is designed to replicate various real-world plate configurations by applying a series of steps to mimic the low-resolution and noisy characteristics of real images. We evaluated the proposed method using a real-world dataset of 783 real car plate images. Our proposed approach which leverages synthetic data samples demonstrated more than 40% reduction in Character Error Rate (CER) achieving performance comparable to a transformer-based Optical Character Recognition (OCR) model fine-tuned with real images. This highlights the utility of synthetic data in OCR tasks, greatly reducing the need for laborious hand-labeled real-world datasets. The code and datasets used in this work are available in the Github repository: https://github.com/DRAGON20-3/Plate_Reading.