Leave the Image Alone

[!NOTE] Pilot Study A pilot preprint is in preparation. The verified stats reports and figure scripts are public in the repository.

Overview

A controlled pilot: 30 CORD-v2 receipts, four preprocessing conditions, four vision-language models, three repeats, for 1,440 transcriptions in total. The classical OCR reflex of binarizing, normalizing, and upscaling before recognition is counterproductive for legible documents read by modern VLMs.

Key Findings

• Raw wins everywhere. Pooled mean character error rate runs raw 0.297, downscale 0.374, 2x upscale 0.728, Otsu binarize 1.460, and the ordering holds for every model tested.
• Binarization is the worst offender. Paired +1.163 CER (95% CI [0.512, 2.390], d = 0.468, p = 0.004).
• Degradation amplifies non-determinism. Within-cell SD grows from 0.051 (raw) to 1.060 (binarized) at temperature 0, showing hosted-model instability is input-dependent.
• A structured-field recall check reproduces the ranking without the CER metric, and the LLM judge auto-calibrates to kappa = 0.70 against hand scoring with zero human labels.

---

Read the Full Paper PDF here. · Paper repository on GitHub

@misc{keough2026leavetheimagealone,
  title  = {Leave the Image Alone},
  author = {Keough, Patrick S.},
  year   = {2026},
  note   = {Pilot Study},
  url    = {https://pskeough.github.io/research/leave-the-image-alone/}
}