OCR Scanned Checks: Amounts, Payees and the MICR Line

OCR can read the printed parts of scanned checks: the account holder's name, the bank, the check number and a typed payee or amount. A person has to confirm everything written by pen, every digit of the MICR line, and whatever sits under a bank stamp. Treat the converted text as a finding aid and the image as the record.

Can OCR read a scanned check?

In part. A check is a small page that mixes clean print, special bank type, pen entries, a patterned background and stamps. Recognition does well on the first of those and gets less dependable with each of the others, so the useful answer goes field by field.

Test with your own three checks

Before you convert a box of checks, try three that differ in kind: a computer-printed business check, a personal check written by pen, and the worst image in the set. Schema Law's free plan takes one document a day of up to three pages, so the three fit in one PDF, and the pricing page shows the page limits for a full set. The site says recognition is strongest on clean print and weaker on handwriting, and that nothing marks a passage as uncertain. For a set written entirely by pen, the post on handwritten checks describes a three-check trial that grades the hands.

Run that test on a personal check and on a cashier's check from a financial institution. A check is a negotiable instrument, so you confirm the payee and both amounts, then the endorsement on the back, including any restriction the depositor wrote. The borrower name on a loan disbursement check gets the same character-by-character pass.

What the text is good for

Converted check text earns its place in two ways. It makes a stack searchable, so a search for a payee's printed name or a check number lands on the right page. And it gives you a first draft of a register that a person then corrects. It does not give you figures you can total without reading each one.

What is the MICR line at the bottom of a check, and can OCR read it?

The MICR line is the row of blocky numerals along the bottom edge. OCR can return its digits, since they are printed shapes like any others, but it reads the picture and not the magnetic ink. The special symbols that separate the fields have no keyboard equivalent, so expect stray characters between the numbers.

Federal banking regulation defines the term. Regulation CC at 12 CFR 229.2(vv) says the MICR line means "the numbers, which may include the routing number, account number, check number, check amount, and other information" that are "Printed near the bottom of a check in magnetic ink". Paragraph (dd) of the same section defines the routing number, and covers a number printed on the face of a check in nine-digit form.

What comes out

A fictional line shows one shape the result can take. Suppose the bottom of the check carries a nine-digit routing number, an account number and check number 2291. The converted text might read:

|:0O1234567|: 4455 8O91 77||' 2291

Three things have happened. The bracket-like bank symbols became pipes, colons and apostrophes. Two zeros became the letter O. And the spacing no longer tells you where one field ends and the next begins.

How to check it

1. Count the digits in the routing number. If the definition's nine-digit form is what your check carries and you have eight or ten, a character was dropped or added.
2. Compare the check number in the bottom line with the check number printed in the top corner. They are printed separately, so agreement is a real cross-check.
3. Paragraph (dd) also covers a routing number printed on the face of a check "in fractional form". If your check shows one, compare the two.
4. Read the account number digit by digit against the image. Unless the check repeats it somewhere else, nothing cross-checks it.
5. Replace any letter in a numeric field with the digit the image shows. The post on why OCR gets numbers wrong lists the swaps to expect, such as O for 0 and l for 1.

The account number is sensitive

A converted check puts a full account number into searchable text. That is useful inside the file and a problem in a filing. In federal civil cases, Rule 5.2(a) of the Federal Rules of Civil Procedure says that, unless the court orders otherwise, a filing that contains a financial-account number "may include only" certain parts of the listed identifiers, among them "the last four digits of the financial-account number". The rule has exemptions, and state courts have their own rules, so read the one that governs your filing. The practical point is that the bottom line of every check image carries the number, in the picture and in the text behind it.

Security backgrounds and patterned check stock

Check paper can carry a tint, a pattern or a picture behind the writing. To your eye the background recedes. To recognition it is a second layer of marks competing with the ink, and how much it interferes depends on how the check was scanned.

What the scan mode does to the pattern

• Black and white. Every pixel becomes ink or paper. A mid-toned pattern lands on either side of the line, so parts of the background turn into black speckle and pale pen strokes turn white. Check images that came from a bank in black and white have already been through this, and no later step restores what was dropped.
• Grayscale. The pattern stays lighter than the ink, and the writing stands out from it.
• Color. Blue or black ink separates from a green or tan background by hue as well as by darkness. Color also keeps red and purple bank stamps distinct.

If the paper checks are yours to scan, use grayscale or color. The comparison of grayscale and black and white scanning explains the trade.

Features that look like faults

Look closely before you decide an image is damaged. A signature line made of tiny printed words scans as a gray or broken rule. A border made of fine repeating lines merges into a band. A scenic or pictorial check puts a darker image behind the payee line. None of these is text you need, and all of them can leave odd characters in the output. Delete those characters from a working copy; do not chase them.

Copies of checks

A photocopy of a check can show a word such as VOID or COPY that the original did not appear to carry. The section on VOID patterns and security paper on certified copies explains printed backgrounds that surface when a page is copied.

Printed and handwritten payees, amounts and memo lines

Typed or computer-printed entries on a check read like any other print. Pen entries do not come with that assurance. Verify the payee, both amounts and the date on every check you rely on, and treat the memo line as something you read from the image.

The two amounts

A check states its amount twice, once in figures and once in words. Read both, and record both in your register, in two columns, exactly as the check shows them. A register that holds only the figure cannot show a conflict. Which amount controls when they disagree is a legal question for the lawyer; the post on handwritten checks quotes the Uniform Commercial Code text on that point under numeric amount and written amount.

On a typed or computer-printed check both amounts are print, so a disagreement in the converted text may come from conversion and not from the check. Look at the image before you record one.

Trouble on the amount lines

Pen entries

Whether software reads a pen-written payee depends on the hand, the pen and the image, and no general statement predicts your set. The guide to what OCR can do with handwritten notes sets out what to expect, and the post on OCR for handwritten checks goes through the pen fields one at a time. For a check, the working rule is short: if the payee, date or memo was written by pen, a person types it from the image.

A memo line deserves the effort. It may hold an invoice number, a loan number, "gift", "loan" or a name, and it sits at the bottom edge where a tight crop can cut it. Confirm that the whole line is in the image before you record it as blank.

A verification order that saves time

1. Check number, against the bottom line.
2. Amount in figures, then amount in words.
3. Date, with the year read last and on its own.
4. Payee, spelled as written.
5. Memo.
6. Whether there is a signature, and whose name is printed on the account. Reading a signature as a name is a judgment; record what you can see.

Endorsements, bank stamps and the back of a check

The back of a check holds the endorsement and the marks each bank added in processing. They overlap, they run in more than one direction, and stamp ink varies from firm to faint. Expect fragments from recognition, and read the back from the image.

First make sure you have the back

A set of check images with fronts only is half a record. Count fronts and backs. If you are scanning paper checks, capture both sides; the post on what duplex scanning is covers two-sided capture and how back sides go missing. If the bank supplied the images, look at how each item is laid out: front above back, front beside back, or backs on separate pages.

Some items in a file are not original checks. Regulation CC defines a substitute check at 12 CFR 229.2(aaa) as a paper reproduction of an original check that, among other things, "Contains an image of the front and back of the original check". Measure the pictures on such an item against an ordinary check. If they are smaller, each letter has fewer pixels, so expect less from recognition and do more by eye.

What is on the back, and how it converts

Rotate the image when you read it. An endorsement written across the short end of the check sits at a right angle to the front. If your converter is given a page of backs that are all sideways, rotate the page before conversion and test whether the stamps read better.

Stamps that overlap each other or sit on top of a signature are the hardest case. The post on whether OCR can read stamps over text explains what scan mode does to overlapping ink.

Why the back matters to the file

The endorsement shows who took the money and where it went. A check payable to one person and deposited into an account with a different number is the kind of fact a register should surface. Give the back its own columns, and do not leave them empty because the text was thin.

Check image pages printed several to a sheet in a bank statement

A statement's check image pages shrink each check to a fraction of its size and print several on one sheet. Recognition has few pixels per letter to work with, so treat these pages as pictures to read. The useful output is a check register you build by hand, tied to the statement's own list of cleared checks.

The transaction list is a table, which is a different conversion job. The guide to bank statements in Excel for family law covers turning those tables into rows and reconciling them.

Why the thumbnails read poorly

A sheet scanned at 300 dots per inch gives each full-size letter a fixed budget of pixels. Print the same check at a third of its width and each letter gets a third of that budget across and a third down. The post on the best DPI for OCR works through the arithmetic. Enlarging the image afterward adds no detail. If the thumbnails cannot be read by a person at high zoom, ask for full-size copies of the checks that matter, front and back.

Columns for a check register

Build the register as its own worksheet, one row per check. The pen-field columns, and the way to mark a blank, illegible or cut-off entry, are set out in the post on handwritten checks under a check register table built from scanned handwritten checks. A register built from statement thumbnails adds the columns that tie each image to the bank's own list.

The statement list and the images check each other. A check number in the list with no thumbnail, or a thumbnail with no row in the list, goes on a list of questions.

A worked example

Tomas Reyes is preparing a tracing schedule for a fictional client, Halvorsen. The production holds twelve monthly statements from a fictional bank, each ending with two pages of check thumbnails.

1. He converts the statements and exports the transaction tables to Excel, one of the export formats listed on the home page.
2. He filters the rows to those with a check number and copies them to a new sheet. That gives him the number, cleared date and amount for 87 checks.
3. He sorts by check number and marks the gaps in the sequence. Numbers 1432 to 1436 never cleared in the period, which he notes as a question, not a finding.
4. He opens the thumbnail pages at 300% and types payee and memo for each check, adding the image page number.
5. For nine checks the thumbnail cannot be read. He lists them by number, date and amount for a request for full-size copies.
6. He compares the amount he reads on each image with the statement amount. One differs by a transposed digit, which turns out to be his own typing.

The register he hands over says, in a header row, that payee and memo were read by eye from reduced images, and names the nine checks still to come.