Upload a meter photo, run local AI detection, and review the extracted reading.
Your AI detection result will appear in this workspace.
Run an analysis to inspect every detected digit.
The Water Meter Reading System MVP was developed collaboratively by three Year 3 students.
Water Meter Reading System project team member.
Water Meter Reading System project team member.
Water Meter Reading System project team member.
Computer vision for detecting water meters, locating their reading windows, recognizing digits, and safely flagging uncertain readings for human review.
Aggregate object-detection results from the final clean YOLO12n model on the held-out 52-image combat test set.
Of the objects predicted, 95.8% were correct.
Of the labeled objects, 95.8% were found.
Detection accuracy at 50% box overlap.
Stricter score averaged across multiple overlaps.
Select a plot to open the original full-resolution evidence.
Interpretation note: These are object-detection validation metrics, not full meter-reading string accuracy. A complete reading also depends on finding the window, filtering digits, ordering them correctly, and applying the unknown rule.
The model detects objects first, then the application filters duplicates and marks uncertain digits for review.
Lower values can recover faint digits but may add false boxes. Higher values show fewer, stronger detections.
Recommended: 10%Controls how strongly overlapping predictions are removed. It does not change what digit the model guessed.
Recommended: 45%A retained digit below this value becomes ?. Its raw guess remains visible so a person can review it.
Recommended: 40%Private by designYour image never leaves the device.
Custom YOLO modelTrained for meter, window and digits.
Browser inferenceONNX Runtime Web, no paid server.