OCR

Bank statement OCR, with the uncertainty left visible

OCR turns pixels into characters. That is only half of a bank statement conversion. The harder part is deciding which characters form a date, which form an amount, and which belong to a page header that should never become a transaction.

FlowParse uses OCR only where a PDF page has no usable text layer, keeps word-level confidence, and flags the rows that depend on a low-confidence read. You get the speed of automation without pretending that a blurry scan is as reliable as a digital statement.

  • Page-level detection decides whether OCR is needed at all.
  • 300 DPI rendering with Tesseract and English language data.
  • Low-confidence pages flag their rows and force a low overall confidence.
  • OCR output is capped at medium confidence, even when the page looks clean.

No card required. 3 processed pages free every month.

How it works

  1. 1

    Check the text layer

    FlowParse first tries to read embedded text. Pages with enough usable text skip OCR entirely, which is faster and more accurate.

  2. 2

    Render and recognise image-only pages

    A page without a text layer is rendered at 300 DPI and processed with OCR. Word positions and confidence are retained for the parser.

  3. 3

    Parse, validate and review

    The parser rebuilds rows from word geometry, reconciles balances, and marks low-confidence rows for review before export.

What OCR does with a bank statement

OCR analyses an image and returns characters with bounding boxes. In a statement, those characters include transaction dates, descriptions, references, debit and credit amounts, running balances, page headers and legal footers.

The parser then groups words into lines, finds the table's column anchors, and assigns each token to a field. Without that second step, OCR output is just a wall of text with no reliable transaction boundaries.

OCR versus a text-layer PDF

Why digital statements are handled differently
Digital PDFScanned PDF
Text sourceEmbedded charactersPixels recognised by OCR
SpeedFastSlower: rendering and recognition per page
CertaintyHigh when the text layer is cleanDepends on scan quality
Typical problemsUnusual layouts, split columnsBlur, skew, decimal confusion
FlowParse handlingDirect extractionOCR plus confidence flags

How confidence is reported

Every OCR word carries a confidence score. FlowParse aggregates it per page. A page with a mean confidence below 65% is marked low confidence: every row from that page is flagged and the statement's confidence is forced to low.

Any statement that uses OCR is capped at medium confidence overall. That is deliberate. A clean-looking scan can still misread a single digit, and the difference between 1,480.00 and 1,460.00 matters more than the visual quality of the page.

What OCR cannot fix

When extraction is not reliable, FlowParse marks the document as needing review or failed and reports the source lines it could not resolve. Manual entry remains available, which is better than a plausible-looking wrong ledger.

  • Handwritten statements, annotations and signatures are not reliable OCR targets.
  • Heavily compressed scans lose the thin strokes that separate 3 from 8 and 1 from 7.
  • Pages photographed at an angle need deskewing that OCR alone does not perform.
  • A missing page cannot be reconstructed from the remaining pages.
  • Statements with a mixed language are limited by the installed English OCR data.

Privacy and retention

Statements are uploaded to a private storage bucket under your user id. The processor downloads a statement with service-role credentials, extracts it, and writes results back through a service-role-only function. The browser never receives the service-role key.

OCR runs on your statement in the processor service; it does not send pages to a third-party OCR API. You can delete a statement from the workspace, which removes its database record and attempts to remove the stored PDF.

When bank statement OCR helps

Legacy statements

The only available copy is a scan from an old filing cabinet or archive.

Client uploads

Clients photograph or scan statements instead of downloading them from a bank.

Mixed batches

Some pages are digital and others are scanned, and you need one consistent ledger.

Audit preparation

You need the transactions as data, but the source only exists as images.

Questions

Does every page go through OCR?

No. FlowParse reads the embedded text layer first and only renders pages that do not contain enough usable text. Digital pages are extracted directly.

What OCR engine and language are used?

Tesseract 5 with English language data, running inside the FlowParse processor. Additional languages require installing the corresponding Tesseract data files.

Can OCR read a photo of a statement?

It can process a PDF built from a photo, but results depend on lighting, angle and resolution. A flat 300 DPI scan is much more reliable. FlowParse accepts PDFs only.

How do I know which rows came from OCR?

OCR pages are marked in the extraction report, confidence is shown for the statement, and low-confidence pages flag every row they produced with review reasons.

Is OCR data sent to a third party?

No. Recognition runs in the FlowParse processor using a local Tesseract installation. Statement pages are not forwarded to an external OCR service.

Convert your first statement free

The Free plan includes 3 processed pages per month. Upload a PDF, review the ledger, and export Excel or CSV when it looks right.