Related Experiment Video
Updated: Feb 28, 2026
04:23
A Swin Transformer-Based Model for Thyroid Nodule Detection in Ultrasound Images
Published on: April 21, 2023
2.4K
A Comparative Study of OCR Architectures for Korean License Plate Recognition: CNN-RNN-Based Models and
1Department of Smart ICT Convergence Engineering, Seoul National University of Science and Technology, Seoul 01811, Republic of Korea.
Sensors (Basel, Switzerland)
|February 27, 2026
Summary
Comparing optical character recognition (OCR) for Korean license plates, this study finds decoder performance varies by dataset and feature quality. Transformer models show high overhead, suggesting integrated systems are key for robust automatic license plate recognition (ALPR).
Area of Science:
- Computer Vision
- Machine Learning
- Artificial Intelligence
Background:
- Automatic License Plate Recognition (ALPR) systems increasingly use Transformer decoders, but the source of performance gains (sequence modeling vs. feature representation) is unclear.
- Existing research often lacks controlled comparisons of Optical Character Recognition (OCR) architectures under identical detection conditions.
Purpose of the Study:
- To systematically compare OCR architectures for Korean license plate recognition.
- To isolate the impact of sequence decoders versus backbone features on ALPR performance.
- To evaluate OCR performance on both static and sequential datasets across different hardware platforms.
Main Methods:
- A unified YOLOv12-based detector was used to provide consistent input to various OCR configurations.
- Evaluated OCR models included CNN with Attention-LSTM and MobileNetV3 with Transformer decoders.
- A controlled ablation study fixed the ResNet-18 backbone while varying only the sequence decoder.
Main Results:
- Sequence decoder effectiveness is dataset-dependent, influenced by feature quality and region-of-interest (ROI) stability.
- Tracking-induced error accumulation significantly impacts OCR performance in sequential recognition.
- Transformer-based OCR models incur substantial computational and memory overhead, hindering real-time deployment on embedded systems.
Conclusions:
- Isolated optimization of OCR architectures is insufficient for robust ALPR.
- Effective Korean license plate recognition necessitates a holistic approach considering detection, tracking, and recognition.
- Dataset characteristics and hardware constraints critically influence the choice of OCR architecture for ALPR.