Prepare an Image for OCR

How it works

Four operations, in a fixed order, with no model anywhere in the chain. The output is a deterministic function of your upload.

First the image is resampled with lanczos filtering to about four megapixels. Then it is sharpened. Then its contrast is raised. Then it is quantized to four tone levels with dithering switched off.

Each of those addresses a different documented failure mode, and the order matters as much as the steps.

Recognition fails in a small number of ways

Text recognition is not a general vision problem. It is a narrow one, and it breaks along predictable lines.

The first is glyph size. Engines are trained on characters at a certain scale and have an effective floor below which accuracy does not degrade gracefully — it collapses. Text photographed small, or scanned at a low setting, sits under that floor and returns nothing usable.

The second is edge definition. Recognition depends on the boundary between ink and paper being decisive. Soft edges from a slightly out-of-focus capture, from compression, or from resampling make every letter's outline ambiguous, and similar characters start swapping.

The third is separation. If ink is dark gray and paper is light gray, the two distributions overlap and no reliable division exists between them.

The fourth is false marks. Paper grain, speckle, dust, and print showing through from the reverse of the sheet all look like small dark things, which is exactly what a character is.

What each step does

Resampling to four megapixels addresses the first failure. This adds no information — interpolation cannot recover detail that was never captured — but it presents the detail that exists at a size the engine was built to read. That distinction matters: a fundamentally unreadable scan stays unreadable, while a readable one that was merely too small becomes readable.

Sharpening addresses the second. Enlargement is inherently softening, so the sharpen restores the abruptness at the ink boundary that the interpolation just smoothed away. It is applied immediately after the resize for that reason.

Raising contrast addresses the third. It pulls the ink and paper distributions apart before anything tries to divide them.

Quantizing to four levels addresses the fourth. Every pixel is forced onto one of a handful of values, so the continuous range where paper texture and faint bleed live is simply removed. Dithering is disabled deliberately — dithering preserves apparent tone by scattering dots, and scattered dots are precisely the artifact that would be read as marks.

This is destructive, and it should be

Worth stating plainly, because it is a reasonable thing to be surprised by.

The output of this page is a worse photograph than the input. Tone is gone, texture is gone, anything that depended on a gradient is flattened. If you print it, it will look harsh.

That is the intent. The result is not for a person to look at — it is an intermediate file for a machine, and every property removed was removed because it makes recognition harder rather than easier.

So keep your original. This is a working copy for one specific purpose, not a replacement for the scan.

Measure it rather than trusting it

The last thing worth saying is that this is a set of reasonable defaults, not a guarantee.

Recognition accuracy depends on your engine, the language, the typeface, and the condition of the source, and it is entirely possible for a particular document to score worse after conditioning than before — most often when the tone reduction takes a light stroke with it.

The check is quick. Run your engine on the original and on the conditioned copy and compare the output. Because nothing here is random, the comparison is stable and the conclusion you draw holds for every document like that one.

Publication gate

This page ships once the workflow has been run against a low-resolution scan, a faded photocopy, a page with print showing through from the reverse, and a page mixing text with a photograph, with recognition accuracy measured against the untreated original in each case.