Preprocessing now tries progressively heavier variants per receipt
(plain -> high-pass -> edge-enhance -> both), stopping as soon as a
cheap LLM-free arithmetic check on the OCR text looks plausible, so
the extra processing only applies to receipts that need it. A flat
version of the same heavier processing regressed the full batch by
helping hard receipts but hurting easy ones; the adaptive version
improves date-matching 37/52 -> 40/52 with no regression (see
testing/2026_10_04_gimp_highpass_nlfilter_experiment.md).
Also adds test_ocr.sh and test_single.sh for checking a single
receipt's OCR text or full pipeline result without a batch run.
Self-hosted, verifiable pipeline for extracting structured data (station,
date/time, price per litre, litres, discount, total) from German fuel
receipt photos: grayscale + mild bilateral filter + contrast stretch,
Tesseract OCR, then LLM structuring via a small local model (Qwen 2.5,
served by Ollama). Each step is independently runnable and inspectable
through numbered scripts against a shared tooling container.
Ollama runs as its own long-lived container (0_run_ollama.sh), separate
from the stateless per-receipt pipeline steps, so the model stays loaded
in memory between runs. Includes a containerized ground-truth batch test
(test_batch.sh) that reproduces the full pipeline over the test receipts
and reports per-field accuracy.
Targets eventual deployment on a Raspberry Pi 4 k3s cluster.
ml
matthias