TITLE:
Unified Cross-Domain Adaptation for License Plate Recognition in Adverse and Multilingual Environments
AUTHORS:
Bertrand Tahte, Geh Wilson Ejuh, Thierry Noulamo, Alain Djimeli, Armand Tchanque Tchakountio, Josiane Tanguebou Kengne
KEYWORDS:
License Plate Recognition, Cross-Domain Adaptation, Multilingual OCR, Oriented Object Detection, Transformer Architecture, Domain Generalization, YOLOv8-OBB, CBAM, Test-Time Adaptation
JOURNAL NAME:
Journal of Computer and Communications,
Vol.13 No.11,
November
27,
2025
ABSTRACT: License Plate Recognition (LPR) systems face significant challenges when deployed in real-world environments that differ from their training data, due to variations in angle, resolution, lighting, plate syntax, and language. This paper presents a unified, robust, and multilingual LPR framework designed to operate effectively in both intra-domain and cross-domain scenarios. The proposed architecture comprises three key modules: (i) an oriented license plate detector based on YOLOv8 with angle regression, enabling accurate detection under oblique or rotated views; (ii) a Transformer-based recognition head enhanced with CBAM and constrained decoding using region-specific syntax; and (iii) a lightweight yet effective domain adaptation module leveraging Maximum Mean Discrepancy (MMD), adversarial training, and test-time self-supervised fine-tuning. Extensive experiments across real-world and synthetic datasets (CCPD, AOLP, UFPR, PKU, LP-Synth) demonstrate that our model outperforms existing state-of-the-art methods, especially in multilingual and cross-domain conditions. Ablation studies further confirm the importance of each architectural component. Our framework opens promising directions for deploying LPR systems in globally diverse and dynamic environments.