OCR Low Quality PDF Scans: A Triage Checklist
By Schema LawPublished Updated
Before you OCR a low-quality PDF, sort its pages into three piles: readable as they are, worth rescanning, and beyond recovery. Convert the first pile, get better copies for the second, and write down the gaps in the third. That triage takes minutes and stops bad text from passing as good.
What counts as a low-quality scan?
A low-quality scan is one where the picture of the page has lost detail that recognition depends on. The letters are too small in pixels, too faint, too thick, tilted, cut off, or mixed with marks that are not text. If you have to squint at 200% zoom, expect the software to struggle too.
A useful working test has two parts:
- Zoom to 200% and read the smallest print on the page. If you can read it without effort, the body text is safe.
- Look at closed letters. Check a lower-case "e", "a" and "o". If the hole in the middle has filled with black, strokes are bleeding together. If the letter has gaps in its outline, strokes are dropping out. Both cause wrong characters.
Resolution is the defect people ask about most, and it is the hardest to see at normal zoom. As one reference point, the documentation for the open-source Tesseract engine, on improving the quality of its output, says it "works best on images which have a DPI of at least 300 dpi". Other engines have their own tolerances, but the direction is the same: fewer dots per letter means more guessing.
Quality also varies inside one file. A 300-page production can hold 280 clean pages and 20 that came off a fax machine in 1998. Judge pages, not files.
A 300-page set is often an electronic discovery production, and the bad pages are the ones a motion to compel or a sanctions letter will quote. You mark a protective order's stamp and any page a person cannot read, so the log shows where the text is too weak to search.
A triage checklist before you run OCR
Page through the PDF once before converting it and mark each problem page against the list below. The aim is to know, before any text exists, which pages you will trust, which you will check line by line, and which need a better copy.
| Check | How to spot it | What it does to the text | What to do |
|---|---|---|---|
| Small or coarse print | Letters look blocky at 200% | Similar characters swap: 1 and l, 5 and S, rn and m | Rescan at a higher resolution if you hold the paper |
| Faint print | Grey letters, broken strokes | Dropped letters and whole dropped lines | Rescan darker, or ask for a better copy |
| Heavy, bleeding print | Closed letters filled in | "e" read as "c" or "o", words run together | Find an earlier-generation copy |
| Tilted page | Lines slope across the screen | Lines merge or break in odd places | See the guide to OCR on a crooked scan |
| Cut-off edges | Text runs into the page edge | Missing first or last words on each line | Rescan or request the full page; cleanup cannot restore it |
| Speckle and background grain | Dots between the lines | False full stops, commas and accents | Check punctuation and decimals |
| Stains, highlighter, folds | Blotches or bars over words | Words under the mark lost or garbled | Compare those lines by hand |
| Stamps or stickers on text | Ink over print | Gaps or merged words | See whether OCR can read stamps over text |
| Fax artifacts | Jagged letters, header line, streaks | Broad errors in small type | See how to OCR a faxed document |
| Phone photo | Curved lines, shadow, uneven focus | Errors grow toward the edges | See how to OCR a photo of a document |
Then give each problem page one of three labels.
- Convert and spot check. Minor defects away from the text that matters. Most pages land here.
- Convert and proofread. Defects that touch names, dates, amounts or operative clauses. Convert the page, then read it against the image.
- Get a better copy. Text a person cannot read with confidence. Software will not do better than you can, and its guess will look as neat as its correct readings.
Write the page numbers down as you go. A list such as "pp. 14 to 19 faxed, p. 44 stained, p. 102 cut off at right" takes two minutes to make and is the start of the record described at the end of this post.
Check the file size while you are there. Colour scans at high resolution are large. Schema Law accepts PDF, PNG and JPEG uploads of up to 20 MB, and each plan sets a page limit for one document, shown on the pricing page. The guide to batch OCR for discovery documents explains how to split a file without losing order.
Is it better to rescan or to OCR what you have?
Rescan when you hold the paper, the defect came from the scanner, and the page matters. OCR what you have when the paper is gone, the defect is on the paper itself, or the page only needs to be findable. A second scan cannot improve on its source.
The first question is where the damage happened. Hold the paper beside the screen if you can.
| Situation | Rescan? | Why |
|---|---|---|
| Paper is clean, scan is faint, coarse or tilted | Yes | The scanner lost the detail, and a second pass gets it back |
| Paper was scanned in black and white and has coloured stamps, highlighter or stains | Yes, in colour or greyscale | The marks and the print can be told apart again |
| Paper itself is a faint photocopy or fax printout | No | A new scan faithfully copies a bad page |
| Page is boilerplate, a cover sheet or a blank form | Not worth it | Being roughly searchable is enough |
| Page holds a figure, date or clause you will quote | Yes, or request a better copy | One wrong character changes the meaning |
Rescan by exception: convert everything, then go back for the pages that need a better copy. The guide to digitizing law firm paper files covers resolution and colour mode, so those settings are not repeated here.
Two cautions apply.
Keep the file you were given. If the PDF came from another party, a court or an agency, store it untouched and work on a duplicate. A rescan or a cleaned-up version is your working copy, never a replacement.
Do not expect a tool to rescue an unreadable page. People often ask what the best OCR for bad scans is. Tools differ by defect, so the only test that means anything is your own worst pages run through the tools you are considering. The post on the best OCR for lawyers sets out how to run that trial fairly. Schema Law's free plan takes one document a day of up to three pages, so you can upload three bad pages and read the parsed result against the originals before deciding anything.
Noise, speckle and coffee stains
Noise is any mark on the scan that is not part of the document: specks, grain, stains, shadows, punch holes. Small marks get read as punctuation. Large marks hide the words under them. Each does a different kind of harm, so each needs a different check.
Speckle
Speckle is a scatter of small dots, usually from a dirty scanner glass, rough paper or a copier low on toner. A dot on the baseline beside a number looks like a full stop or a comma, and "1,250" and "1.250" differ by one dot. Check every figure on a speckled page.
Stains and uneven backgrounds
A coffee ring is darker than the paper and lighter than the ink. A person reads through it. Software first has to decide which pixels are ink, and a stain makes that harder. The Tesseract documentation is candid about it: converting an image to black and white "can be suboptimal, particularly if the page background is of uneven darkness". Yellowed paper and show-through from the reverse side cause the same trouble. In practice the words under a stain may vanish, or the stain's edge may be read as a line of stray characters.
Borders, holes and fold lines
Dark borders appear when the page was smaller than the scanner bed or the lid was open. The same Tesseract page notes that such borders "can be erroneously picked up as extra characters". Punch holes come out as "O" or "0" in the margin. A fold line can be read as a row of dashes or can split a line of text in two.
Cleaning the image first is rarely worth it. Image editors and some scanning programs offer despeckle, background removal and contrast controls. A despeckle filter that removes dots may remove real full stops and the dots on "i" and "j". A strong contrast setting can erase faint print altogether.
If you clean a page, follow three rules. Work on a copy. Compare the cleaned image with the original before you convert it. Note in your file that the working copy was altered. For most pages it is quicker to convert the page as it is and proofread the affected lines.
Photocopies of photocopies
Each time a page is photocopied, it loses detail that no later step can restore. Strokes thicken or thin, small print closes up, edges creep and the page tilts a little more. By the fourth or fifth generation the characters themselves have changed shape, and OCR reads the shapes it is given.
The signs of a late-generation copy are easy to learn:
- Filled or broken letters. Heavy copies close the holes in "e", "a", "o" and "8". Light copies break thin strokes, so "h" loses its arch and "m" comes apart. The Tesseract documentation makes the same point about character weight, saying that bold or thin characters "may impact the recognition of details and reduce recognition accuracy".
- Touching letters. "rn" merges into "m", "cl" into "d", "vv" into "w".
- Lost fine detail. Underlines merge with descenders, superscript footnote numbers vanish, and the dot of a decimal point disappears into the digits around it.
The fix is nearly always upstream. Ask who holds an earlier generation, and ask for that. The client may have the signed original. A recorder, clerk or agency may hold the official copy of a filed or recorded document. An earlier filing in the same case may carry a cleaner version of an exhibit than the latest one does.
One request for a better copy of a named page often costs less than an hour of proofreading, and it improves every later use of that page. When no better copy exists, convert what you have, treat the page as label two or three from the checklist, and expect to type the key passages yourself. The post on why OCR output is full of errors lists the character swaps to look for.
Recording what could not be read
When part of a page cannot be read, write that fact down where the next person will find it. A short log, kept with the converted files, turns an invisible weakness in the text into a known one. Without it, a search that finds nothing looks the same as a search of pages that were never readable.
If someone searches a set for a name and gets no hits on pages 14 to 19, they will conclude the name is not there. If those pages are a smudged fax, the right conclusion is that nobody knows yet.
Keep the log as a spreadsheet or a table in a memo. These columns are enough:
| File | Page | Where on the page | Problem | What was done | Checked by |
|---|---|---|---|---|---|
| Smith production vol. 2 | 14 to 19 | Whole pages | Fax of a photocopy, small type | Converted; names and dates proofread; body not relied on | JR |
| Smith production vol. 2 | 44 | Lower third | Stain over payment table | Figures typed by hand from the image | JR |
| Lease, 1994 | 7 | Right margin | Last word of each line cut off | Better copy requested from client | MK |
A few habits make the log worth having:
- Mark gaps in the text itself. Where you correct or transcribe a passage in a Word export, put "[illegible]" for what nobody can read. Never fill in what the clause probably said.
- Say who read it. A passage typed by a person from the image is more reliable than a machine reading of a damaged page, and the reader should be able to tell which is which.
- Tell searchers. Put the list of unreliable page ranges at the top of any search memo, so "no hits" is read with the right caveat.
- Keep the image with the text. A searchable PDF keeps the page picture, so anyone who doubts a passage can look at the page itself. Schema Law lets you export the same upload again in another format, so you can keep a searchable PDF for the file and a Word copy for working notes.