A Multilingual License Plate Recognition Framework Integrating Super-Resolution and Domain Adaptation

Abstract

The deployment of license plate recognition systems in a multinational environment still poses a challenge due to differences in plate designs, script types, visual representations, and image acquisition conditions. Substantial performance degradation of single-domain trained models has been reported when deployed in unseen domains. An integrated multilingual license plate recognition framework is presented in this work, aimed at improving model performance under diverse acquisition conditions. A dual super-resolution enhancement strategy combining Real-Enhanced Super-Resolution Generative Adversarial Network (Real-ESRGAN) and Enhanced Deep Super-Resolution (EDSR), a You Only Look Once version 8 (YOLOv8)-based plate detection module, and a Convolutional Neural Network (CNN)-based multilingual Optical Character Recognition (OCR) system supporting Latin, Arabic, and Chinese scripts are incorporated within the framework. A Cycle-Consistent Generative Adversarial Network (CycleGAN)-based unpaired image translation technique is also employed during the training process. Experiments conducted on European, Chinese, and Moroccan datasets show consistent recognition performance across three geographically distinct regional subsets under diverse acquisition conditions. The results indicate that integrating image enhancement and domain adaptation within a multilingual license plate recognition pipeline contributes to improved performance.