Analysis-by-synthesis decoding of severely degraded license plates: render candidate plate strings through a physical camera forward model (optics → sensor → ISP → codec), score hypotheses by likelihood against the observed pixels, decode exactly via a character-position trellis, and fuse video frames by log-likelihood summation.
Primary scientific goal: the ideal-observer performance bound for the task — Bayes-optimal recognition rate as a function of degradation — and statistical-efficiency analysis of existing recognizers against it.
See generative-plate-decoding-research-memo.md for the full research memo (v2).
Requires uv.
uv syncsrc/lrlpr/— library code (renderer, forward model, trellis decoder, fusion)tests/— pytest suitedocs/— research notes and phase write-upsdata/— datasets (gitignored; see docs for access instructions)