Transformer-Based Detection Model for Number Recognition on Electric kWh Meters

Manual recording of analog kWh meters frequently results in user complaints due to discrepancies between recorded and actual electricity usage. These issues stem from the continued reliance on manual data collection. This study proposes a model that automatically detects and extracts numerical values from kWh electricity meters using the Detection Transformer (DETR) for object detection and EasyOCR for optical character recognition (OCR). The model was developed using the Machine Learning Life Cycle (MLLC) methodology, comprising data acquisition, preprocessing, modeling, evaluation, and deployment. Evaluation using the Mean Average Precision (mAP) metric yielded a score of 96.83%, demonstrating high object detection accuracy. The trained model was integrated into a simple web application built with the Flask framework. While the model performed well on high-quality images, its effectiveness declined on low-quality images, such as blurry or distant captures. This study highlights the potential of DETR for object detection and OCR-based text extraction in analog meter reading, while also identifying challenges in handling suboptimal image conditions for future improvements