Everyone says OCR is a timesaver, but on 120 crumpled cafe receipts yesterday, Tesseract plus cleanup took longer than just tabbing through fields in Google Sheets and still only hit about 93% accuracy… If you’ve got a setup where OCR really beats manual on messy inputs, what’s the trick — preprocessing steps, templates, or a tool I should try instead?
I only got OCR to beat manual when I used a receipt-specific model and “ironed” the image first. Auto dewarp/deskew + CLAHE, then Azure’s prebuilt receipt model instead of raw Tesseract bumped me from about 93% to about 98% and about halved time on crumpled slips (Receipt data extraction - Document Intelligence - Foundry Tools | Microsoft Learn). If you need to stay local, whitelist digits/currency and add a vendor wordlist — does your set skew to one chain, @OP?
Messy cafe slips are , but what sped me up was narrowing scope: OCR only date/merchant/total with a char whitelist and regex, and auto-queue anything below 0.90 confidence for manual. That took me from “93% accuracy” to about 99% on about 150 receipts and beat tabbing in Sheets; line items I still enter by hand. Piggybacking on @riverledger, try Tesseract --psm 6 with tessedit_char_whitelist=0123456789./ and a quick total-vs-line sum sanity check — does that sound workable for your 120?
Quick win: run two Tesseract passes (psm 6 and 11) and majority‑vote just the total/date/merchant, then auto‑accept only when subtotal+tax≈total; anything that fails that checksum dumps to a manual queue. Building on @sward, this cut my hand checks by about 60% because the math acts like a bouncer — no match, no entry.
Same boat here. What finally tipped it for me on about 100 crumpled thermal slips was switching to tessdata_fast (GitHub - tesseract-ocr/tessdata_fast: Fast integer versions of trained LSTM models) and pre-inverting + Sauvola thresholding before OCR; have you tried that on your 120 receipts? It’s a small accuracy trade‑off versus best, but overall throughput beat tabbing in Sheets.
Have you tried a prebuilt receipt model instead of generic OCR? Azure Form Recognizer’s Receipt (Receipt data extraction - Document Intelligence - Foundry Tools | Microsoft Learn) plus a simple ‘auto-reject if subtotal+tax!=total’ and a small merchant dictionary ended up quicker than keying by hand on wrinkly café slips, with the caveat of per‑page cost and data leaving your machine.
Piggybacking on @bravementor: what sped me up wasn’t new OCR, it was a merchant-first flow — fuzzy-match the first text line to bucket by cafe, then run tiny vendor-specific regex templates so you only review outliers. On about 150 crumpled slips that beat manual and pushed past the “93%” you hit. Would a quick merchant classifier plus 3–5 templates fit your 120?
and a quick total-vs-line sum sanity check — does that sound workable for your 120?