Remove Shadows from a Photographed Document

How it works

Your image is scaled so its total area is around two megapixels, keeping the original proportions. Then a single morphological operation does the work, followed by an inversion and a contrast adjustment.

There is no model here. No diffusion, no sampler, no prompt, no seed. The graph is four image operations in a row, and the output is a deterministic function of the input — the same file always produces the same result, because there is nothing in the chain that could produce anything else.

Separating by size, not by darkness

This is the idea worth understanding, because it explains everything the tool does well and everything it loses.

A photographed page contains two signals mixed together. The illumination varies slowly and broadly: a gradient from the window, a soft falloff towards a corner, the wide dark region under your own shadow. The ink varies quickly and narrowly: a letter stroke is a few pixels wide and goes from paper to full ink and back within that distance.

Brightness cannot tell them apart. A letter in a lit region is far brighter than blank paper in the shadow, which is why thresholding a photographed page produces a black corner and a white corner and nothing useful in between.

Spatial scale can tell them apart, easily. So the operation asks a different question: what would this page look like with the writing removed?

It answers by taking a morphological closing with a thirty one pixel kernel. Closing eliminates dark features smaller than the kernel, so every letter is filled in with the paper tone immediately around it — including that paper's local shading. What comes back is an estimate of your sheet with no ink on it, shadow and all.

Subtract that estimate from the original and the shading cancels, because it is present in both. What remains is only the part that was too small to survive the closing: the text. Invert it, and you have dark writing on clean white paper, with the lighting gone.

What the kernel size decides

The thirty one pixel figure is the one meaningful parameter, and it sets a boundary.

Anything narrower than the kernel is ink. Anything wider is background. That is the whole classification, and it is why the method is so robust for body text — a letter stroke at this working size is a few pixels across, nowhere near the threshold.

It is also why large solid marks vanish. A display heading with thick strokes, a solid black logo, a filled cell in a table, a photograph printed on the page: to the closing these are indistinguishable from a dark patch of paper, so they are absorbed into the background estimate and subtracted away with the shadow.

That is a real limitation and it is not tunable from the page. Documents made of text come out excellently. Documents that are half graphic do not, and it is better to know which of those you have before uploading.

When to use this instead of the AI pages

Most of this site repaints images with a generative model. For a document that is the wrong tool, and it is worth being blunt about why.

A repaint reconstructs the picture, including the characters in it. The model does not read your document; it produces something that looks like a document, and a digit whose shape was ambiguous becomes a confident digit that may not be the one that was printed. Nothing flags it. The result looks cleaner than the original and can be quietly wrong.

For an invoice, a form, a prescription, a meter reading, or anything that will be relied on later, that risk is unacceptable and unnecessary. This page cannot make that mistake, because it has no capacity to generate. Every mark in the output came from a mark in your upload.

Publication gate

This page ships once the workflow has been run against a page with a gradient shadow, a page with a hard shadow edge, a document containing a thick heading and a filled table, and a page of faint pencil writing, with the output compared to the source character by character.