The processing pipeline
1
Capture
You start the camera or upload a photo or video file. When using the live camera, the app captures frames continuously at up to approximately 50 frames per second and feeds them into the detection pipeline.
2
Plate detection
A YOLOv9 AI model scans each frame for license plate regions. It analyzes the entire frame in a single pass and draws a bounding box around any plate it finds.
3
Text recognition (OCR)
Each detected plate region is cropped out of the frame and passed to a MobileViT v2 OCR model. This model reads the characters on the plate and assigns a confidence score to each one.
4
Quality check
Each result is scored against four criteria: character length, mean confidence, minimum per-character confidence, and plate format. Only results with a combined quality score of 0.7 or higher are kept. Results that fall below this threshold are discarded silently.
5
Confirmation window
A plate must be detected consistently before it is added to your history. In standard mode, the same plate must appear for 3 continuous seconds. If the detection confidence is very high (mean confidence ≥ 0.8), this window shortens to 1 second.
6
Grouping
Once confirmed, new detections are compared against plates already in your history. If two plate readings are at least 80% similar (measured by Levenshtein string distance), they are grouped together as the same plate. This prevents minor OCR variations — such as a single misread character — from creating duplicate entries.
All processing runs in a dedicated Web Worker thread, separate from the browser’s main UI thread. This keeps the interface smooth and responsive even while the models are actively running.
Plate quality validation rules
Before a plate is saved to your history, it must pass all four of these criteria:
Each criterion carries a different weight in the combined score: mean confidence counts for 30%, format and minimum character confidence each count for 25%, and length counts for 20%. A plate must reach 0.7 overall to be accepted.