Vision & OCR
The plate pipeline and its published Phase 14 benchmark.
Pipeline
A classical OpenCV plate-region detector (blackhat → Sobel-x → Otsu → close → aspect-ratio contour filter) finds candidate plates; EasyOCR 1.7.2 reads them on the CPU. Camera sources can be USB, RTSP or a video file. A read counts as trusted at confidence ≥ 0.5.
Phase 14 evaluation
Measured on a seeded, synthetic dataset — 1,250 images: 50 plate texts under 23 conditions, plus 100 scenes with no plate. It characterises the pipeline; it is not a production accuracy claim.
| Result | Value |
|---|---|
| Exact plate reads (misses count as wrong) | 77.4% (890 / 1,150) |
| Character accuracy on detected plates | 93.76% |
| False plates on empty scenes | 0 / 100 |
| Precision of trusted reads | 85.0% |
| Condition | Exact |
| --- | --- |
| Clean, lighting, sensor noise | 100% |
| Rotation and perspective | 98.5% |
| Gaussian and motion blur | 75.0% |
| Distance | 50.5% |
| Partial occlusion | 33.3% |
docs/vision/phase14-evaluation.md.