OCR

How to Convert Scanned Bank Statements Using OCR

How OCR turns scanned bank statements into usable transaction data, what affects accuracy, and how to review low-confidence rows before exporting to Excel or CSV.

By FlowParsePublished 9 min read

A scanned statement is a picture. The PDF contains no text layer, so selecting it returns nothing and a normal parser sees a blank page. OCR is the step that turns those pixels into characters the parser can work with.

This guide explains where OCR fits, how to prepare a scan, and why a responsible conversion reports confidence instead of pretending every recognised digit is correct.

When a statement needs OCR

There are two kinds of PDF. A digital statement exported by a bank contains text and font information; selecting a word highlights it. A scanned statement contains an image per page, sometimes with an invisible OCR layer added by the scanner.

The first kind can be parsed directly. The second needs recognition before parsing. Mixed documents exist too: a digital statement with a scanned appendix, or a bundle where some pages were printed and re-scanned.

How OCR works on a bank statement

Recognition is only the first half. A page of OCR text without row reconstruction is still not a spreadsheet, and a table read in the wrong column order produces plausible-looking values in the wrong fields.

  • The page is rendered to an image at a chosen resolution, typically 300 DPI.
  • The OCR engine detects text regions and lines, then recognises characters.
  • Each word is returned with a bounding box and a confidence score.
  • The parser groups words into rows, finds the table columns and assigns fields.
  • The reconciliation checks the result against the statement's balances.

What affects OCR quality

Scan factors and their effect
FactorEffect
ResolutionLow DPI loses thin strokes; 300 DPI is a sensible target
SkewTilted text makes columns drift and rows merge
ContrastFaded print reduces character confidence
CompressionHeavy JPEG artefacts blur digits and decimal points
Page geometryMulti-column layouts can be read in the wrong order
LanguageOCR data must include the statement's language

Preparing a better scan

  • Scan at 300 DPI or higher in greyscale or colour.
  • Keep the page straight; most scanner software has a deskew option.
  • Include the whole page, including headers and totals.
  • Avoid photographing a screen; moiré patterns confuse recognition.
  • If the bank offers a digital PDF, use it instead of a scan.

Confidence and review

An OCR engine returns a confidence score per word. A mean page confidence below 65% is a strong signal that the page needs a human look. FlowParse marks those pages, flags every row they produced and forces the statement's confidence to low.

Any statement that relies on OCR is capped at medium confidence overall. That is not a bug; it is a statement about the source. A clean-looking scan can still misread one digit, and a single wrong digit can break a reconciliation.

What OCR cannot fix

When the result is unreliable, the correct outcome is a clear failure or a review flag, not an invented transaction. FlowParse lists unresolved source lines so they can be entered manually.

  • Handwritten annotations and signatures.
  • Torn, creased or partially obscured pages.
  • A missing page: the transactions on it cannot be recovered.
  • Statements in a language for which no OCR data is installed.
  • A scan so compressed that decimal points disappear.

From scan to workbook

Once the rows are extracted, the export is the same as for a digital statement: XLSX for analysis or CSV for an import. The difference is the review note and confidence attached to the OCR pages. Start with Convert a scanned bank statement to Excel or read more about bank statement OCR.

Frequently asked questions

Is OCR accurate enough for bookkeeping?

It can be, with review. Digital statements are more reliable. For scanned pages, check the confidence flags and reconcile the totals before posting anything.

Can I OCR a photo of a statement?

You can convert the photo to a PDF and upload it, but a flat scan produces better results. FlowParse accepts PDFs, not image files directly.

Does OCR run locally or in the cloud?

In FlowParse's processor service using a local Tesseract installation. Pages are not sent to an external OCR API.

Why are all rows on a page flagged?

The page's mean OCR confidence was below the threshold. That flags every row from the page so the reviewer knows the whole page is uncertain, not just one line.

Read less, convert more

Put the guide into practice on a real statement. The Free plan includes 3 processed pages per month, with no card required.