<![CDATA[The implementation of Automatic License Plate Recognition (ALPR) in Indonesia faces challenges arising from visual degradation and anomalies in real-world datasets, including annotation errors, data leakage, and imbalanced distributions of regional license plate codes. This study aims to develop a precise and adaptive end-to-end ALPR system for Indonesian vehicle license plates by optimizing both license plate detection and character recognition architectures. A Data-Centric AI approach was employed to improve dataset quality through bootstrap relabeling and the targeted synthesis of 2,773 images to address regional distribution imbalances. The proposed system adopts a two-stage processing pipeline, consisting of license plate localization using YOLOv11m and text recognition through a comparative evaluation of PaddleOCR, Tesseract, EasyOCR, and TrOCR-base. Specifically, TrOCR was optimized using a two-stage fine-tuning strategy (K4 Two-Stage) that combines synthetic data during pretraining with real-world data during subsequent fine-tuning. The results demonstrate that the proposed framework effectively mitigates dataset anomalies. The YOLOv11m model achieved a license plate detection mAP@50 of 97.05%. Among the recognition models, TrOCR K4 achieved the best performance, with an Exact Match (EM) accuracy of 86.21% and a Character Error Rate (CER) of 3.03%. End-to-end inference evaluation achieved an overall exact-match accuracy of 87.07%, with a low inference speed of 2.84 FPS (approximately 206 ms per license plate) on an NVIDIA L4 GPU. Overall, the integration of Data-Centric AI, YOLOv11, and TrOCR produced a highly accurate and efficient Indonesian license plate recognition system, demonstrating strong potential for real-time vehicle monitoring applications.]]>
