Enhancing Water Meter Digit Recognition Using YOLO-OCR
Abstrak
This study presents a deep learning-based framework for automatic water meter digit recognition by combining You Only Look Once version 8 (YOLOv8) for digit region detection and PP-OCRv4 for Optical Character Recognition (OCR). The model was developed using a dataset of 3,250 annotated water meter images collected under various lighting and visibility conditions, including ideal and non-ideal scenarios. The detection module was fine-tuned using different optimizers, with Adam achieving the highest detection accuracy (F1-score, precision, and recall all at 100% and a Mean Average Precision at Intersection over Union (IoU) thresholds ranging from 0.50 to 0.90 (mAP50–90) of 0.95831). The recognition model was trained on both the full dataset and a subset of 1,000 non-ideal images. Interestingly, the best performance was observed in the model trained solely on non-ideal data using AdamW, achieving a recognition accuracy of 90.90% and a Character Error Rate (CER) of 0.0151 in end-to-end testing across 55 benchmark images. The system demonstrated near real-time inference, averaging 0.5 s per image. These results validate the robustness and practical applicability of the proposed YOLO-OCR pipeline and highlight the advantage of training on visually challenging data to improve robustness in real-world deployment conditions.
Identifier
- DOI: https://doi.org/10.48084/etasr.17291
- OpenAlex: https://openalex.org/W7163747137