Best DPI for OCR: What Resolution Means on a Scanned Page

The best DPI for OCR on ordinary typed pages is 300 dots per inch. Move up to 400 or 600 when a page carries small print, and expect little from higher settings on normal type. The number counts the pixels a scanner records along each inch of paper, which decides how many pixels draw each letter.

That last point is the one worth holding on to. Recognition software does not care about the setting on the scanner's control panel. It cares how many pixels it has to work with across the height of a letter. Resolution and type size together produce that figure, so a setting that suits a twelve-point pleading can fail on the footnotes of the same page.

What is the best DPI for OCR?

The best DPI for OCR is 300 for typed pages in ordinary sizes, which means roughly ten-point type and larger. Use 400 for pages with footnotes or fine print, and 600 for very small or reduced type. Below 300, letters start to lose the details that tell one from another.

Published guidance clusters around the same figure, with some spread:

• The documentation for the open-source engine Tesseract, on improving the quality of its output, says the engine works best on images with a DPI of at least 300.
• Amazon's best practices for Textract ask for a high quality image, "ideally at least 150 DPI". That is a floor for one service, not a target for a scanning project.
• The federal regulation on digitizing permanent paper records, 36 CFR 1236.50, sets a minimum of 300 ppi for modern textual paper records. It binds federal agencies, not law offices, but it shows where a careful institution drew the line.

A starting point by kind of page:

You will see both "dpi" and "ppi" used. Strictly, a scanner samples pixels per inch and a printer lays down dots per inch. Scanner menus and vendor pages commonly use the two as the same thing for scanned images, and this post does too.

Resolution is one of several settings that shape a scan. Color mode matters almost as much, and the comparison of grayscale and black and white scanning explains that choice. For duplex, paper preparation and file format, see the full list of scanner settings for a digitizing project.

What does 300 dpi mean on a sheet of letter paper?

At 300 dpi, a letter-size sheet of 8.5 by 11 inches becomes an image 2,550 pixels wide and 3,300 pixels tall, about 8.4 million pixels in all. Each inch of paper, in each direction, is cut into 300 samples, and each sample becomes one pixel.

The same sheet at other settings:

Notice that the total grows with the square of the setting. Doubling the dpi doubles the pixels along each edge, so the page holds four times as many.

From page pixels to letter pixels

Type is measured in points, and a point is one seventy-second of an inch. That gives a simple formula:

pixels of height for a line of type = point size × dpi ÷ 72

Twelve-point type at 300 dpi gets 12 × 300 ÷ 72, which is 50 pixels. That 50 is the full body of the type, from the top of a tall letter to the bottom of a "g" or "y". A lowercase letter with no ascender, such as "a", "e" or "o", is roughly half of that in most typefaces, so about 25 pixels tall.

These pixel counts are this post's own arithmetic from the point size and the scanner setting. They are not a threshold published for any OCR engine, and typefaces vary, so read them as rough sizes.

The arithmetic shows why the setting matters. What separates "c" from "e" is a thin crossbar. What separates "rn" from "m" is a narrow gap. In a letter 25 pixels tall, each of those features is a few pixels of the image. At 150 dpi the same "e" is about 12 pixels tall, and the crossbar and the gap shrink to a pixel or less.

Judge 300 dpi against the smallest type you care about on one transcript. That may be a citation, a statute section, or a party name in a footnote. Tiny type at that setting can leave the letter shapes too coarse. Search for the citation after the scan, and raise the setting if the hit fails.

Why small print needs more dots per inch

Small print needs more dots per inch because the scanner spends the same number of pixels on every inch of paper, whatever is printed there. A six-point letter covers half the height of a twelve-point one, so at the same setting it gets half the pixels.

Height in pixels is point size times dpi, divided by 72, because a point is one seventy-second of an inch. Worked for common sizes:

The cells are pixels of height for the full body of the type, worked out with that formula. They are arithmetic, not figures taken from any engine's documentation. Read the table diagonally. Eight-point type at 300 dpi gets the same 33 pixels as twelve-point type at 200 dpi, a setting nobody recommends for body text. Six-point type needs 600 dpi to reach what twelve-point type gets at 300.

A rule you can apply at the scanner: find the smallest type on the page that you need to be right, and divide 3,600 by its point size. The answer is the dpi that gives it 50 pixels, the height twelve-point type gets at 300 dpi. Fifty is this post's yardstick, taken from that arithmetic, and no engine publishes it as a requirement. For 8 point the answer is 450, so choose 400 or 600. For 9 point it is 400.

Where small type hides in a legal file

• Footnotes. A brief set in twelve point often has footnotes in ten or smaller, and that is where the citations are.
• Condensed transcripts. A deposition printed four pages to a sheet shrinks every line to about half size. A normal scan of that sheet is a low-resolution scan of each page on it.
• Terms on the back of a form. Invoices, bills of lading and credit applications often print their terms in pale six-point type.
• Schedules and riders. Insurance declarations and loan schedules pack figures into narrow columns.
• Superscripts and subscripts. Footnote markers and exponents are smaller than the line they sit on.
• Photocopies made at reduced size. A legal-size page copied down to letter size has lost about a fifth of its type height before you scan it.

A worked example

Suppose you are scanning a fictional equipment lease between Harlow Freight and Castell Leasing. The body is eleven-point type. The payment schedule on the last page is seven point. At 300 dpi the body gets 46 pixels a line and reads cleanly. The schedule gets 29, and suppose the output shows "$1,875.00" in one row as "$1,675.00". Rescanning that one page at 600 dpi gives the schedule 58 pixels a line. You do not need to rescan the body. Scan by the smallest type that matters on each page, not by the document.

Does scanning above 300 dpi improve OCR?

Scanning above 300 dpi improves OCR when the type is small or the strokes are fine. On ordinary ten to twelve-point print it adds pixels the software does not need. The cost is certain either way: at 600 dpi each page carries four times the pixels of a 300 dpi page.

Once every letter has enough pixels to show its shape, more pixels describe the same shape in finer steps. The engine learns nothing new about whether the mark is an "e". What the extra resolution does capture is paper grain, toner spatter and the weave of a photocopy's gray background, none of which helps.

The costs are practical ones:

• File size. Compressed files do not grow by exactly four times, but they grow a great deal. A long exhibit can pass an upload or e-filing limit. Schema Law accepts uploads up to 20 MB, and a 600 dpi grayscale scan reaches that in far fewer pages than a 300 dpi one.
• Time. Feeders run slower at high resolution, and every later step, from saving to uploading to opening the file, takes longer.
• Handling. Large files stall in email, document management systems and review platforms.

Higher resolution does help in a few defined cases:

If you are setting a standard for a large project, test before you commit. Take one typical page and one with the smallest print in the set. Scan each at 300, 400 and 600 dpi with everything else unchanged. Run all six through the tool you plan to use and compare the hardest lines against the paper. Choose the lowest setting at which the small print comes out right.

DPI in a phone photo or a file with no resolution recorded

A phone photo has no true DPI, because a camera records a fixed number of pixels however much paper is in the frame. What matters is effective resolution: the number of pixels that fall across the page, divided by the page's size in inches. The same applies to screenshots and to image files with no resolution recorded.

Many image files carry a resolution label of 72 or 96. Ignore it. It is a default written by the camera or the software, and it says nothing about how the picture was taken. Work the figure out from the pixels:

1. Find the image's dimensions in pixels. On most computers, the file's properties or the image viewer's information panel show them.
2. Estimate how much of the frame the page fills. If there is a wide border of desk, measure or estimate the page's share.
3. Divide the pixels across the page by the paper's width in inches.

Some examples on letter paper:

Two lessons follow. First, a photo taken from close enough that the page fills the frame can match a 300 dpi scan in pixel count, though blur, shadow and perspective are separate problems, covered in the post on reading a phone photo of a document. Second, a screenshot of a full page is almost always too coarse. If you must capture from a screen, zoom in and take the page in sections, or better, get the file itself.

Messaging apps and email clients often shrink photos on sending. A picture that left the phone at 4,000 pixels wide may arrive at 1,600. Ask the client to send the photo at "actual size" or "original", or to share it through a file link.

Scanned dpi versus the dpi a PDF was saved or downsampled at

The dpi a page was scanned at and the dpi of the PDF in front of you can differ, because resolution is easy to throw away after scanning. A PDF holds each page image at whatever pixel count it was last saved with. If a step in between reduced the pixels, the original setting no longer matters.

The usual places it is lost:

• "Reduce file size" and "optimize" commands. These commonly resample page images to a lower pixel count and compress them harder.
• Scan-to-email presets. A copier's "compact" or "small file" profile may scan at a low setting, or scan high and save low, to keep attachments under a mail limit.
• Printing a scanned PDF to a new PDF. The print path can re-render each page at the driver's resolution and compression.
• Fax. A page that passed through a fax line was sampled coarsely on the way, and the post on reading a faxed document covers that case.
• Portals and messaging. Upload forms and chat tools sometimes recompress images to save space.

Amazon's guidance on input documents for Textract makes the same point from the vendor's side: if a document is already in a supported format, do not convert or downsample it before uploading.

Checking what a PDF really holds

You can estimate the effective resolution of a PDF page without special software:

1. Zoom to 400% on small type. If letter edges turn to visible square blocks and thin strokes vanish, the image is coarse.
2. Look at the size per page. Divide the file size by the page count. Compare it with a page you scan yourself at 300 dpi in the same color mode. A file that is a small fraction of that size per page has probably been downsampled or compressed hard.
3. Export one page as an image and read its pixel dimensions. Divide the width in pixels by the page width in inches. A letter page that exports at 1,275 pixels wide is a 150 dpi image, whatever the scanner was set to. Make sure the export uses the image's own resolution and not a setting you chose in the export dialog.

Enlarging does not bring detail back

Resampling an image upward, from 150 to 300 dpi, creates new pixels by averaging their neighbors. The file reports a higher figure and contains no more information. Tesseract's documentation notes that resizing an image may help that engine, which is about giving it letters at the size it expects, not about recovering detail. If the crossbar of an "e" was never recorded, no enlargement draws it.

So the order of preference is fixed. Use the first-generation scan if it still exists, which means keeping the full-size master whenever you make a reduced copy for email. Ask the sender for the original file if you received a shrunken one. Rescan the paper if you hold it. Only then work with the coarse file, and the post on low quality PDFs helps you decide whether to rescan or use the file you have.

When you are unsure whether a file is good enough, test a page. Schema Law's free plan reads one document a day of up to three pages, so you can upload the page with the smallest print and compare the result with the original. The pricing page lists the limits for longer documents.