Benchmark methodology

How we measure OCR accuracy.

Accuracy claims are only as good as the methodology behind them. Every figure on the License Plate OCR page comes from a single public, reproducible benchmark run — here is exactly how it was measured.

Results — July 2026, no location hints
100% plate detection 96.4% exact plate reads 98.7% character accuracy 98.8% on high-confidence reads median processing under 50 ms

The benchmark

We evaluate on the OpenALPR public US benchmark — 222 real-world images of US plates in the wild: highways, parking lots, varied angles, lighting and distances. It is the industry-standard set commercial ALPR vendors quote, which makes our numbers directly comparable. The run was measured in July 2026 with no location hints: every image was scored cold, with no state_suggestions prior — the hardest configuration, and the one your accuracy can only improve on.

Scoring convention: O and 0

The letter O and the digit 0 are scored as equivalent, the standard industry convention — they are visually identical on most US plate dies, and many DMVs treat them as interchangeable. Scored strictly, with O and 0 counted as distinct characters, exact plate reads are 95.0%. Both figures are reported by the evaluation harness on every run.

Ground-truth corrections

Two of the benchmark's 222 labels are wrong. After manual inspection of the original images, we corrected them — and rather than silently editing the benchmark, the corrections are published alongside our evaluation harness so any run is reproducible and auditable:

wts-lg-000085.jpg 5DJ0529 BDJ0529 (the leading glyph is definitively a B)
wts-lg-000159.jpg 59E21P1 59321P1 (the third glyph is a 3, not an E)

Evaluation-only — never trained on

The benchmark is used strictly for evaluation. Its images are excluded from all training data, so these figures measure how the system generalizes to plates it has never seen — not how well it memorized a test set.