Abstract:
Automatic License Plate Recognition (ALPR) has become a widely used computer vision application applied in various
fields. However, traditional ALPR methods face challenges as they require high-quality images from a fixed-angle camera to
produce accurate results in recognizing License Plate (LP) characters. Weather conditions and unfavorable LP angles can lead to
low-resolution images, causing inaccuracies in LP character recognition by LPR systems. There is a need to improve LPR systems to
adapt to a variety of captured image conditions, particularly those with low resolution. To address this issue, past researchers have
developed Super-resolution (SR) models capable of generating high-resolution images from low-resolution counterparts. In this
study, we enhance the LPR system by incorporating SR, aiming to improve character recognition. The study comprises two phases:
License Plate Detection (LPD) and License Plate Recognition (LPR). In the LPD phase, we utilize state-of-the-art object detection
models, including Faster R-CNN using detectron2 and YOLOv8. Our detection models perform well, especially YOLOv8, which
achieves 93% accuracy in both train and validation datasets, slightly decreasing to 90% in the test dataset. This outperforms Faster RCNN,
which achieves 71%, 71%, and 74%, respectively. In the recognition phase, we employ two approaches: Tesseract-OCR alone
and Tesseract-OCR with SRGAN. The end-to-end pipeline achieves a Character Error Rate (CER) of 53.9% and a Levenshtein
distance of 3.6% without SR-GAN. When SR-GAN is applied as a preprocessing step, the CER is reduced to 51.7%, and the
Levenshtein distance is reduced to 3.5%. This highlights the effectiveness of SR-GAN in enhancing image quality and, consequently,
improving the performance of OCR engines. The insights gained from this study can contribute to the development of robust license
plate recognition systems for real-world deployment.
