BankStatementDesk

How to check OCR accuracy in a scanned statement before export

A scanned statement can look perfect and still put the wrong number into your workbook. The picture you see and the recognized characters underneath it can disagree. To check OCR accuracy, compare the extracted fields against the visible source, give extra attention to signs and separators, and use independent totals to find rows needing a second look. Keep a marked working copy so you can explain why a particular value changed.

In this guide
  1. Establish a baseline before recognition
  2. Read fields in pairs
  3. Worked example with a missing decimal point
  4. Work through suspect words and plausible mistakes
  5. Turn review findings into a usable export
  6. Decide when a fresh scan is worth the effort

Suppose you have twelve pages from a paper archive and a deadline for entering supplier payments. Reading every word aloud is a slow starting point. Begin by establishing what the result must contain. For this task that might be transaction date, payee, payment, deposit, and closing balance. A mistake in an advertisement near the footer can wait. A misplaced decimal point in a payment cannot. Design the review around the fields that will actually be reused.

Establish a baseline before recognition

Choose a representative page and record a few values directly from the image, including one that looks difficult to read. Include a negative amount, an identifier containing a zero, a description wrapping onto a second line, and a balance near the page edge if those features occur. Record the printed page label and a nearby date so you can find each value again. Avoid using an earlier extract as the answer key for a new extraction.

Run recognition on a working copy with the language appropriate to the text. When the software offers orientation or image correction, check that the page is upright and that a pale last column has survived the preparation. Apply settings to a small sample before processing the entire packet. A setting that makes bold headings wonderfully clear can make a faint decimal point hard to distinguish. Judge the result using the row values you recorded, alongside the image itself.

Keep the original and the searchable copy distinct

Name the copies so that another reviewer can identify the original image, the searchable working file, and the exported data. A simple sequence such as “supplier-march-original,” “supplier-march-searchable,” and “supplier-march-reviewed” is enough when it matches your filing practices. If you later repeat recognition with different settings, set the candidate copy aside until comparison is complete. Overwriting your only reference can remove the easiest way to settle a disagreement.

Read fields in pairs

Work from image to text for part of the review and from text back to image for another part. Reading the printed payee and then locating it in the extract helps find omissions. Starting with an unexpected number in the extract and locating the corresponding printed line helps find duplicates or orphaned continuation rows. These are different checks. Performing both helps you catch errors that a straight top-to-bottom reading can conceal.

For amounts, read the digits around each separator, then read the sign and the column label. A faint comma can move a value between interpretations in different regional formats. Parentheses may carry the negative meaning for an entire amount. A narrow figure one can look like a lowercase letter in a reference number. Avoid a universal find-and-replace for a character such as O. It can be wrong in an amount and entirely correct in a supplier name.

Isolate character errors from column errors

If the original shows a debit of 58.20 and the recognized text also contains 58.20, the character recognition may be fine. If the exported spreadsheet shows 58.20 under deposits, the problem arose when the table was reconstructed or interpreted. Correcting the OCR word alone will not fix that assignment. Compare the text layer, the position on the source page, and the preview column before changing underlying data.

Worked example with a missing decimal point

Consider a fictional statement segment. The balance before two purchases is 846.20. A purchase of 18.40 is followed by one of 26.75. The last printed balance is 801.05. The searchable copy reads the first purchase as 1840 and the second as 26.75. Visually, the printed first purchase looks plausible until you zoom in. The blank space between eight and four contains a small decimal point almost touching the paper texture.

Compare the movement shown by the printed balances. The opening 846.20 minus the closing 801.05 gives a reduction of 45.15. The printed purchases add to 45.15. The recognized purchases in their current form add to 1866.75. This arithmetic tells you to inspect the amounts. It does not independently identify which one is incorrect. Check both against the image, correct the first row to 18.40, and repeat the calculation using the revised values.

Record what changed and where the evidence came from

For an internal review note, write “Page four, stationery purchase, amount originally extracted as 1840, image shows 18.40.” A colleague can now repeat the check without having to infer which transaction you meant. If a value remains illegible, mark it unresolved and obtain a better source where possible. Inserting the figure needed to force a balancing total replaces evidence with an assumption, particularly when multiple rows may have been misread.

Work through suspect words and plausible mistakes

Some OCR software highlights text it considers uncertain. Use that queue to find weak recognition areas, especially repeated marks at the edge of the document. Adobe describes the comparison of recognized words and the underlying image in its guide to correcting OCR errors. After reviewing those suggestions, still check your transaction amounts and identifiers. A perfectly plausible recognized word can be the wrong one for the document.

Consider a supplier code printed as B018. If recognition produces B0I8, a reviewer may not notice until the import fails to match the supplier list. Search the original working document for other examples of the same code and compare its typeface. If you have an authorized reference list, consult it as a second clue. Keep the value unresolved if the page and the reference list disagree. A familiar supplier can have several legitimate subaccounts.

Check page boundaries and mixed layouts

Scans of bound statements often have an inner margin that darkens on alternating pages. A sample from page one might say little about the rightmost balance column on page two. Test both odd and even pages if the physical scan shows this pattern. Likewise, sample old and new styles separately when a packet spans a template change. If only one page is faint, replacing that scan may accomplish more than tuning recognition for every other page.

Turn review findings into a usable export

Once the text layer is good enough for extraction, open the bank statement to Excel converter and inspect the preview against the source. For a general-purpose table, use the PDF to CSV converter. Neither conversion supplies missing characters from an image-only page. If nothing is extracted, return to the recognized copy and test whether the transaction text can actually be selected and copied.

Review the count of real entries separately from the number of extracted lines. A statement with seventeen printed transactions may also have a header, a balance-forward line, and three wrapped descriptions. The worksheet may therefore show more than seventeen lines without containing an extra transaction. Conversely, two missing transactions and two extra headings can make the totals look deceptively tidy. Establish how you identify a genuine transaction in this particular document before comparing counts.

Preserve uncertainty during a handoff

A useful handoff separates verified and unresolved work. Give the next reviewer the source page references and the reason each row remains open. For example, “Page seven, second transfer, last digit of amount obscured by fold” is more actionable than “check OCR.” Store the original alongside the corrected copy in an approved location. If another department only needs the completed totals, share the amount of source material appropriate for that task and your normal document-handling rules.

Decide when a fresh scan is worth the effort

Recognition cannot recover printed detail that is absent from the image. Look at a missing character in context. If neighboring rows have sharp edges, the source may have a genuine blank or a damaged spot. If the entire column fades toward the page edge, request a flatter, clearer capture of that side. Confirm that the replacement includes the full page, including the final line and the printed page number, before discarding any working material.

For a document with repeated errors in the same field, run one controlled comparison. Recognize the same representative page from the new capture, measure the same sample fields, and compare the number of confirmed corrections. Keep the version that preserves the evidence most clearly. A nicer-looking thumbnail or a smaller file size is a poor reason to choose between two copies whose decimal points and signs must support later calculations.

How much checking is enough?

For a quick personal search, finding the correct page and reading its image may meet your purpose. For entering amounts into a spreadsheet that someone else will use, verify the consequential fields and reconcile them to independent source values where available. When a result affects a decision, leave the degree of review visible. “Amounts checked against each printed entry; descriptions checked where uncertain” says exactly what has happened and what a future reader can rely on.

If two reviewers divide a large packet, give them a shared example of a resolved row and a shared way to name unresolved entries. Have both reviewers independently check one difficult page and compare their findings before splitting the remaining work. Disagreement about whether brackets mean a debit, for instance, is worth resolving before that interpretation reaches dozens of rows. For fields with recurring uncertainty, record the source convention and the example that established it.