Real-Time Water Meter Recognition System for Edge Devices: OCR Studio’s Paper Published in the SPIE proceedings of ICMV 2025

Last year, OCR Studio made its debut at the 18th International Conference on Machine Vision (ICMV 2025) in Paris, France. Among other technologies, our researchers presented a real-time water meter recognition system designed for edge devices. After several months of waiting, we are delighted to share that the corresponding paper has now been published in the electronic proceedings of the International Society for Optics and Photonics (SPIE). Learn more about our new paper and the technology behind it in today’s blog post.

What OCR Studio’s New Paper Is About

In this paper, we consider the problem of recognizing water meters from a photo directly on devices with limited computing power. A feature of the proposed approach is high quality of work without strict requirements on the meter’s position or its scale in the frame. The detection and recognition models are resistant to complex distortions in the photo and allow you to recognize readings at even extreme capture angles.

The paper provides performance measurements of full-frame recognition on mobile and desktop processors, which demonstrate that any mobile device is suitable for meter scanning. In addition, lightweight algorithms help to reduce battery consumption.

Our On-Device Approach to Water Meter Detection and Recognition

Our approach is focused on recognizing meters directly on a mobile device without transferring photos to any external servers. To achieve this, the proposed pipeline consists of only two components – detection of the odometer-type meter display and end-to-end recognition of the numeric reading in the found zone. A lightweight neural network model is used for each task, which makes the system applicable even on low-budget mobile devices.

To localize a meter dial, a neural network was based on the lightweight YOLO-WM (Water Meter) model with 480 thousand trainable parameters. To facilitate zone recognition and eliminate extra steps in the pipeline, YOLO-WM predicts the oriented bounding box of the meter.

Challenges of Water Meter Recognition

In terms of recognition, meters are complicated not only because of image distortions due to challenging capture conditions, but also because of the meter design itself. The digits in analog meters are printed on rotating rollers, which often results in intermediate positions of the numbers. Due to the wide variety of types and designs of meters, the task of determining the position of the comma, separating the readings of cubic meters from liters, is also non-trivial.

To avoid dividing the zone recognition task into digits segmentation and classification steps, we propose a recurrent convolutional neural network for end-to-end dial recognition. The advantage of this approach is that a single model takes into account the scene context, which is important for the determination of the comma position, as well as the correct recognition of digits given the overall design of the meter. The proposed CRNN-WM model uses a lightweight backbone with ResNet blocks and a GRU block with an attention mechanism that takes into consideration the zone context and sequentially builds the recognition result. In total, the model consists of 390 thousand parameters.

Instant Recognition on Edge Devices

In terms of performance, recognition is instantaneous for the user even on the old iPhone 7 (2016), taking just 91 ms per frame. On the newer, but still not the most top-end iPhone 13 (2021), the time is 34 ms. We also tested the performance of the method on AR glasses, for which a computationally efficient pipeline is also critical because of short battery life. Frame recognition on RayNeo X2 AR Glasses, powered by the Qualcomm Snapdragon XR2 chip, takes only 147 ms. The proposed system allows more efficient and faster meter reading without requiring a stable Internet connection.

About OCR Studio

OCR Studio, a developer of optical character recognition solutions, remains committed to a science-driven approach to innovation. Every year, our researchers take part in leading international conferences, where they present our latest advances in document recognition, ID authenticity verification, and machine-readable objects scanning. This ongoing scientific work helps us transform cutting-edge research into practical technologies for real-world use. Find out more about our solution for water meter reading here.