OCR Blurry Scan Pages: Focus, Motion and What to Redo
By Schema LawPublished Updated
OCR can read a blurry scan only while the blur is mild. Letters that are soft but still separate convert with extra errors, and letters that have run into each other do not convert reliably at all. If you hold the paper, rescan the blurred pages. If you do not, convert, proofread and record which pages were weak.
Can OCR read a blurry scan?
Partly. Software reads the letters that are still distinct on a soft page, and guesses at the ones that are not. The guess is the danger: a blurred "8" that comes out as "3" is printed in the same clean type as every correct digit beside it, with nothing to show that the page was hard to read.
The Tesseract project is frank about the limit in its notes on improving the quality of OCR output. The engine cleans up each image before it reads, and the documentation says "there will inevitably be cases where it isn't good enough, which can result in a significant reduction in accuracy". A blurred page is one way to reach that point with any engine, so test your own tool on your own soft pages.
The quiet error shows up on a soft police report or a lab page you will treat as forensic evidence. A quoted statement can look tidy while a time or a badge number changed. Check that one figure from the paper if you still have it.
Three grades of blur
Sort a soft page by looking at one thing: the white gaps inside and between letters.
| Grade | What the gaps look like at 200% | What to expect | What to do |
|---|---|---|---|
| Mild | Edges are fuzzy, but the hole in "e" and the gap between "r" and "n" are still white | Body text converts; small print and figures pick up errors | Convert, then proofread names, dates and amounts |
| Moderate | Holes in "e" and "a" are gray; neighboring letters touch | Wrong letters throughout, merged words, lost punctuation | Rescan if you can; otherwise proofread the whole page |
| Severe | Words are gray bars with bumps | Lines dropped or turned into nonsense | Get a better copy, or type the passage from the paper |
The broader decision between a new scan and the file you already have turns on who holds the paper and where the damage happened. The triage guide for poor scans works through that in its section on whether to rescan or OCR what you have.
Two look-alike problems need a different fix. Print that is sharp but pale is a contrast problem, and the post on OCR for faded text covers it. A page that is short of pixels is a resolution problem, which the post on low-resolution scans takes up.
Is the page blurry or only low in resolution?
Zoom to 400% and look at the edge of one letter. On a low-resolution page the edge is a staircase of hard, square blocks. On a blurred page the edge is a smooth ramp from black through several shades of gray to white, with no hard boundary anywhere. The first lacks pixels. The second has pixels and wasted them.
The two get confused because both make small print hard to read. They have different causes and different cures, so the minute spent telling them apart is repaid at once.
| What you see at 400% | Diagnosis | What fixes it |
|---|---|---|
| Square steps on diagonals; edges of the steps are crisp | Low resolution | A scan with more dots per inch, or the full-size file from the sender |
| Wide gray ramps; curves are smooth but nothing is crisp | Blur | Better contact, a clean scanner, or an earlier copy of the page |
| Large squares, each one smeared | Both at once | A new scan from the paper; neither cure alone is enough |
| Smooth letters in one part of the page, soft in another | Local blur | Find out what lifted or slipped at that spot |
Why more resolution does not cure blur
A page that floated a few millimeters over the glass is out of focus at any setting. Scan it again at twice the resolution and you get four times the pixels holding the same soft letters in more detail. Raising resolution helps only when the scanner was seeing the page sharply and recording too little of it.
The reverse mistake also happens. Someone looks at a soft page, opens the file properties, sees a figure of 300 and concludes that the scan is fine. That figure counts pixels. It says nothing about whether the page was in focus when they were recorded. If the question you have is about the pixel count itself, the post on OCR for a low-resolution scan shows how to measure it and what the numbers mean.
A second test: find something that should be sharp
Look for a feature that was physically present at the scanner: the cut edge of the sheet, a punch hole, a staple, a handwritten note in fresh ink. Compare it with the print.
- Paper edge sharp, print soft. The scanner was in focus. The softness was already on the paper, which points to a photocopy of a photocopy or a fax printout.
- Everything soft, edge included. The capture was out of focus or the page moved.
- Fresh ink sharp, typed text soft. The typed text was copied before the note was added. You are holding a late-generation copy with an original annotation.
That one comparison tells you whether another pass through your own scanner can help at all. Blur in a photograph of a page has its own causes, and the guide to OCR on a photo of a document deals with those.
What makes a flatbed or sheet-fed scan come out blurred?
Four things blur a scanner's picture of a page: the paper was not flat against the glass, the paper moved unevenly through the feeder, the glass or the optics were dirty, or the page was already soft because it is a copy of a copy. Each leaves a different pattern, and the pattern tells you which one you have.
| Cause | Where the blur sits | What it looks like | Fix |
|---|---|---|---|
| Page lifted off a flatbed | One corner, one edge, or along a fold | Sharp where the paper touched, soft where it rose | Flatten the page and close the lid with gentle weight |
| Uneven travel through a feeder | A band across the page, at right angles to the direction of feed | Letters stretched, squashed or smeared for a few lines, then normal | Clean or replace rollers; feed the sheet alone |
| Film or dust on the glass | The whole page, or the same stripe on every page | Low contrast and a gray haze; streaks along the feed direction | Clean the glass, including the narrow strip a feeder scans through |
| Copier generations | The whole page, evenly | Thick, rounded letters with filled holes | No scanner fix; find an earlier copy |
Lifted pages
A flatbed focuses at the surface of its glass. Some models hold focus a little way above it and others lose it within a millimeter or two. You can learn which kind you have in one test: scan a printed sheet with a coin under one corner and see how quickly the print softens toward the coin.
Paper lifts for ordinary reasons. A stapled corner holds the top sheet away from the glass. A page that spent years folded in thirds rises at each crease. A sheet still in a binder or a bound volume curves away at the spine, and the guide to crooked scans covers curved pages from bound volumes. A thick exhibit tab keeps the lid from closing.
Feeder motion
A sheet-fed scanner records one thin line at a time while rollers pull the paper past a fixed sensor. If the paper slips, drags or stalls for an instant, the lines recorded during that instant are spread over the wrong length of paper. The result is a horizontal band in which letters are too tall, too short or smeared. Worn rollers, a creased or torn leading edge, sticky notes, and mixed paper weights in one stack are the things to look at.
Dirty glass and optics
Haze on the glass lowers contrast across the page, which makes soft edges softer. On a sheet-fed scanner the scanning window is a strip about as wide as a pencil, and a fleck of correction fluid or adhesive on it draws a line down every page. A duplex scanner has a window for each side of the sheet, so one dirty window blurs or streaks every second page of the file.
A rescan routine for a blurred page
- Remove staples, clips and sticky notes, and flatten folds by hand.
- Wipe the glass, and the feeder strip if the scanner has one, with a dry lint-free cloth.
- Place the page alone on the flatbed, face down, squared to the corner guide.
- Close the lid fully. For a stiff or curled sheet, rest a hardback book on the lid.
- Scan in grayscale at the resolution your office uses for text, and open the result at 400% before you put the paper away.
- Check the same letter that was soft before. If it is still soft with the page flat and the glass clean, the blur is on the paper.
Does sharpening a blurry page help OCR?
Sometimes a little, and never as much as a clean rescan. A sharpening filter steepens the edges that are still in the image. It cannot rebuild an edge that the blur erased, so two letters that merged into one gray shape stay merged, only with a harder outline.
What a filter can and cannot recover
| Filter | What it does to the image | Can help when | Cannot do |
|---|---|---|---|
| Sharpen or unsharp mask | Raises contrast along existing edges | Blur is mild and letters are still separate | Separate letters that already touch |
| Contrast or levels | Pushes grays toward black or white | Soft letters are also pale | Decide correctly which gray was ink and which was paper |
| Despeckle | Removes small isolated dots | Sharpening has brought up grain | Tell a speck from a full stop or the dot on an "i" |
| Enlarging the image | Adds pixels by estimating between existing ones | Letters are small as well as soft | Add detail the scan never recorded |
Sharpening has costs. It strengthens every edge, including the grain of the paper and the specks on it, and specks beside a figure get read as decimal points and commas. Pushed hard, it draws a pale halo around each stroke, and the halo can cut a thin letter in two. Enlarging is its own subject, covered in the low-resolution post under whether enlarging a low-resolution image helps recognition.
How to test sharpening on your own page
Do not take a filter on trust, and do not judge it by eye. A sharpened page can look better to a person and read worse to software.
- Pick one blurred page and duplicate it.
- Apply the filter to the duplicate at a low strength.
- Convert both versions with the same tool.
- Choose a fixed passage of about ten lines that includes figures, and count the wrong characters in each result against the paper or the image.
- Keep whichever version has fewer errors, and stop there.
If the two counts are close, skip the filter. You then have one fewer altered image to explain.
Keep the altered image apart from the record
A filtered page is a working copy. Store the file you received, or the first scan you made, untouched under its own name. If text from a sharpened image ends up in something you file or serve, your note should say that the image was adjusted before conversion, so nobody mistakes the working copy for the document.
Letters that merge first when a page is soft
Blur closes the smallest white gaps first. That means the holes inside letters, the slivers between thin neighboring strokes, and the space around dots and commas. Those are the places to proofread on a soft page, well before you read the long words, which carry enough shape to survive.
The order in which a soft page fails
| What closes | Typical misreading | Where it matters |
|---|---|---|
| The hole in "e", "a", "o" | "e" read as "c" or "o"; "a" as "o" | Surnames, defined terms |
| The gap between "r" and "n" | "rn" read as "m" | "modern" and "modem", "burn" and "bum" |
| The gap between "c" and "l" | "cl" read as "d" | "clause" becomes "dause" |
| The dot over "i" and the stem beneath it | "i" read as "l" or "1" | Mixed letter and number codes |
| The openings in "3", "5" and "6" | Read as "8" | Amounts, dates, account numbers |
| The space around a decimal point or comma | Point dropped or read as a comma | Every money figure |
| The gap between a colon's two dots | Colon read as a full stop or a single mark | Times, ratios, citations |
Size and weight change the order. Footnotes and fine print close up before body text. Bold type closes before regular, and a condensed typeface before a wide one. A superscript footnote number can vanish into the letter it follows.
A worked example
A fictional invoice from Hartsfield Cold Storage to a client reads, on paper, "Invoice 3186, dated 03/05/2019, balance $8,860.50". The scan is moderately soft. A plausible converted line is "Invoice 3188, dated 08/05/2019, balance $8,880.50". Three digits changed, each from an open shape to "8", and the sentence still looks perfectly sensible. Nothing about the text tells you to doubt it.
So proofread a soft page in this order:
- Money amounts, including the position of the decimal point.
- Dates, digit by digit.
- Account, invoice, case and policy numbers.
- Names of people and companies.
- Section and paragraph numbers.
- The body text, read for sense.
For the character pairs that swap on any imperfect scan, soft or not, see the list of which digits and letters get swapped most often.
Blur on one edge or one page of an otherwise sharp file
Local blur has a local cause, and its position names it. Soft print down one edge means the sheet lifted there. A soft band across the page means the feeder slipped. One soft page in a hundred means that sheet misfed. Every second page soft means one side of a duplex scanner needs cleaning.
Reading the pattern
- Along the inner margin: the page was scanned while still bound or stapled.
- At one corner: a staple, a clip or a dog-eared corner held the sheet up.
- A band at the same height on several pages: a flat spot or dirt on a roller.
- Whole pages at intervals: thicker inserts, such as tabs or card covers, travelling differently through the feeder.
- All odd or all even pages: the scanning window for that side of the sheet.
Finding the bad pages in a long file
Nobody pages through 600 images at 400%. Thumbnails do not help either, because blur disappears when a page is shrunk. Let the converted text point to the pages instead.
- Convert the whole file.
- Search for a common short word, such as "the". Pages with no hits, in a file of running prose, are suspects.
- Compare the amount of text per page. A page with a third of the words of its neighbors, where the image shows a full page of print, was read badly.
- Run a spelling check over the export and note where the flagged words cluster.
- Open only the suspect pages at 400% and note, for each, whether the gaps inside and between letters are still white, gray, or gone.
A spreadsheet makes step 3 quick. Schema Law's homepage describes its CSV export as the document's tables, or its text line by line with page numbers. With the line-by-line form, a count of lines per page takes one formula. CSV is one of the eight export formats on every plan, and the pricing page shows how many pages each plan converts. On any plan you can view the parsed pages on screen and compare a doubtful one with its original before exporting.
Record what you found
Write the page numbers down with the cause and the action: "pp. 212 to 215, soft band mid-page, rescanned from paper" or "p. 88, blurred in production as received, better copy requested". A search that returns nothing on a blurred page has not shown that the term is absent. The list tells the next reader which silences to distrust.