Boundary-Aware Character Parsing for Real-Time Licence Plate Recognition in Degraded Traffic Scenes

Abstract

Automatic licence plate recognition remains difficult in open traffic scenes because plate images are often affected by low illumination, motion blur, occlusion, skew, weather degradation and cluttered backgrounds. This study proposes a boundary-aware character parsing framework for real-time licence plate recognition. The proposed contribution is the Boundary-Aware Character Parsing approach, which utilizes learned transitions between characters based on enhanced token representations via attention mechanisms for grouping tokens adaptively before CTC-based sequence decoding. The resulting boundary map guides adaptive token grouping, after which a connectionist temporal classification decoder recognises the character sequence without requiring fixed character alignment. The framework is designed to reduce segmentation errors in plates with touching, distorted or unevenly spaced characters while maintaining practical inference speed for intelligent transportation, parking management and surveillance applications. Experiments are reported on the CCPD dataset and on the AOLP access-control, law-enforcement and road-patrol subsets, with augmentation for rotation, brightness variation, blur and spatial shifts. Compared with CRNN, LPRNet, ViTSTR, YOLOv8 plus OCR, and DETR plus OCR baselines, the proposed framework achieves 97.2% recognition accuracy, 97.0% precision, 96.6% recall and a 96.8% F1-score, while using 9.5 million parameters, 22.8 GFLOPs and an average inference time of 21.4 ms. Ablation experiments indicate that attention encoding, boundary supervision and attention regularisation each contribute to improved recognition performance, with the full model outperforming the CTC-only baseline by 3.3 percentage points. Additional error analysis shows that low illumination, motion blur, occlusion and extreme skew remain the dominant failure modes. The proposed framework works with the help of a pre-cropped image of the license plate obtained from external detector. This approach does not carry out license plate detection for any vehicle. Instead, it only concentrates on license plate character recognition by predicting the entire sequence of characters from the input image through token grouping and CTC decoding.