Automated Water Meter Reading Through Image Recognition

Water meter readings within the Philippine setting are primarily done manually, which is error-prone and work-intensive. Existing smart solutions for the problem, such as smart water metering, are costly to implement and challenging to scale up as it requires replacing each water meter. The most cost-effective alternative is to utilize image capture and deep learning methods to perform image recognition on water meters. This work created an end-to-end automated water meter reader system with image capture via mobile phone that a human meter reader can use to perform meter readings. This reduces the typical multi-step operation of manual transcription to taking a photo of the meter with a mobile application, with additional features to ensure secure and reliable image capture of the meter in case there is a dispute in the reading. The system consists of a mobile application, a cloud database, an image recognition pipeline, and an automated process that retrieves, reads and logs the result. The system performs well on foreign meters that are similar to the training data but performed comparatively poorly on meters used locally in the Philippines, with a reading accuracy of 91.5% and 75%, respectively for the whole digits of a meter counter.