Enhance & Resolve
A tail-number crop pulled from a ramp photo and knocked down to surveillance quality — 64 pixels wide, heavy sensor noise. At that size the registration is an unreadable smudge, and baseline OCR misreads the tail number. One pass of overcast’s enhance — cleaned up and upscaled — recovers the true registration, confirmed against the full-resolution source.
smudge reads E-RUC
→
cleaned up + upscaled
→
reads EI-RJC
→
source confirms EI-RJC
Before — 64px, surveillance qualityinput
OCR reads E-RUC — a garbled, wrong tail number
After — cleaned up + 2× upscaledenhance
OCR reads EI-RJC — the true registration, recovered (deterministic, no AI)
After — generative super-resreconstructive
OCR reads EI-RJC — crispest to the eye, but the pixels are invented
Source control — full resolutionground truth
OCR reads EI-RJC — CityJet Avro RJ85, confirms the recovery
⚠ Enhancement is reconstructive. The deterministic clean-up (denoise + 2× upscale) invents no detail and read the tail number correctly — confirmed here against the full-resolution source. The generative super-resolution is the crispest to look at but synthesizes pixels: on repeat runs it slips the final letter (EI-RJO / EI-RJQ), so it is a lead to corroborate, never sole proof.