Scan a Document With Your Phone Camera, Free

How it works

1. Take the photos or choose the ones already on the phone. Get the whole sheet in frame with a bit of space around it and do not worry about holding it square. They are opened here on your own device and never uploaded.
2. The edges of each sheet are found for you and shown as a box with four corners. If one corner landed on the edge of a clipboard instead of the paper, drag it where it belongs. The page that comes out is shown next to the photograph while you do it.
3. Pick how it should look, Document being the one that reads as a scan, then Make the PDF. One photo or sixty, they come down as a single file.

What you get

• Finds the sheet in the photograph by itself, and shows you the four corners it chose before anything is committed
• Drag any corner, or the whole box, when it picked the clipboard instead of the paper
• Pulls the page square, so a sheet photographed from one side comes out as a rectangle and not a trapezoid
• Lifts the paper to white by measuring how bright it is in each part of the page, which is what removes the shadow your own hand cast
• Keeps ink the color it is, so a blue signature stays blue and is still recognizable as a signature
• Grayscale and black and white for the times somebody asks for a clean copy, and the untouched photograph for the times it is evidence
• Works out what shape the sheet was from the angle you photographed it at, or put it on letter, legal or A4 yourself
• Says so when the sheet runs off the edge of the picture, because that part of the page was never photographed and cannot be recovered
• Up to sixty pages in one file, each one shown as a thumbnail you can take back out

A photograph of a document is not a document

Everyone has sent one. The lien waiver comes back signed at four o'clock, the office needs it before five, and there is no scanner within forty miles of the job. So the phone comes out, the page gets photographed on the tailgate or the hood or the passenger seat, and it goes out as an attachment.

It is legible. That is not the same as usable. The page is at an angle, so it is a trapezoid rather than a rectangle. Half the frame is truck bed. The paper is gray on one side and blue on the other because half of it is in your shadow and the other half is in the open sky. Print it and the printer lays down every bit of that: the tailgate, the shadow, the lot. Put it in a packet next to a real scan and the difference is not subtle.

The specific thing that gives it away is the shadow, and it is worth saying why. A scanner drags its own light across the sheet, so the paper is the same brightness everywhere and the only dark thing in the image is the writing. A camera has no such luxury. Whatever light is in the yard falls across the page unevenly, and the hand holding the phone puts a gradient right through the middle of it. Everything else about a phone photo is fixable by cropping. That is not.

Which is the whole job here. Find the sheet in the picture, throw away everything that is not the sheet, undo the angle, and then take the lighting out of the paper without taking the ink out with it.

How the sheet gets found

There is a classic way to do this and it is not the way this does it. The classic way looks for edges: run an edge detector across the picture, run a Hough transform over the edges to find straight lines, and work out which four lines make a rectangle. It is a good method on a clean shot and it falls apart the moment the page is on a workbench with four other straight things on it, because a tape measure and the edge of a plank are also excellent straight lines.

Finding the paper turns out to be easier than finding its edges. A sheet of paper in a photograph is the one large bright thing with no color in it. So the picture is shrunk, each pixel is asked whether it is bright and uncolored, the threshold for what counts as bright is worked out from the picture itself rather than fixed, and what comes back is a map of everything that might be paper. Keep the biggest connected patch of that, fill in the holes the writing punched in it, and you have the sheet.

Then the corners get read off the patch directly. The trick that fails here is the obvious one: take the point with the smallest x plus y as the top left, and so on. That works until the page is turned forty degrees, at which point it quietly returns four points along one edge. So the patch's own axes are measured first and the corners are found along those, which is why a sheet photographed sideways comes out as well as one photographed straight.

It was written and scored outside this site before it was put in it, against photographs of a letter sheet whose corners were known to the tenth of a pixel. The worst corner it has ever placed was seven and a half pixels out on a two thousand pixel diagonal, which is under four tenths of one percent, in between fourteen and fifty one milliseconds. And a photograph with no paper in it at all is refused rather than answered, which is the part that matters more, because a tool that is confidently wrong about where the page is is worse than one that asks.

And how it gets pulled flat

Once the four corners are known, straightening the page is arithmetic. There is exactly one projective transform that takes those four points to the four corners of a rectangle, and running the finished page backwards through it says which pixel of the photograph each output pixel came from. Sample those with a bit of blending between neighbors and the trapezoid becomes a rectangle. It is the same mathematics a surveyor uses to rubber sheet an aerial photo onto a map.

The question that is not arithmetic is what shape the rectangle should be. A photograph does not know how many inches anything is, and simply averaging the lengths of the opposite edges is wrong whenever there is any perspective in the shot, because the near edge of a tilted page is genuinely longer in the picture than the far one. Average those and a letter sheet comes out noticeably too wide.

There is a better answer and it is a small piece of geometry. The two pairs of opposite edges of the sheet each meet at a point off in the distance, the two vanishing points, and because the real sheet has square corners, the directions to those two points have to be at right angles in the world. That single fact is enough to recover what the camera's lens was doing, and from that, the true ratio of the sides. It is measured off every photograph here and it lands within about two percent of a letter sheet's real shape, which is close enough that nobody looks at the result and thinks something is off.

It only works when there is some angle to work with. Photograph a page dead square on and the vanishing points run off to infinity, the geometry has nothing to bite on, and it says so and falls back to averaging the edges, which on a square on photograph is exactly right anyway. Either way you can overrule it: tell it the sheet was letter, legal or A4 and that is the paper it goes on.

Taking the shadow off without taking the ink off

This is the part that decides whether the result looks scanned or looks like a photograph somebody brightened. The instinct is to raise the brightness and the contrast until the paper goes white. Do that and the half of the page that was in shadow goes gray while the half in the sun blows out, because one adjustment is being applied to a page that was never evenly lit in the first place.

What works is to divide the picture by its own background rather than subtracting one number from all of it. Every pixel is asked how bright it is compared to the paper right there, not compared to the paper on average. Paper divided by the paper next to it is one, so it goes white everywhere, in the shadow and in the sun alike. Ink divided by the paper next to it is still dark, so it stays.

Which leaves the question of how you know how bright the paper is at a spot that has writing on it. The answer is that paper is brighter than anything written on it, so the background is the brightest thing in any small neighborhood. The picture is shrunk by keeping the largest value in each block rather than the average, which lets a line of type vanish instead of dragging the estimate down with it, and the small picture is then smoothed into a gentle map of the lighting. At an eighth of the size that costs a couple of milliseconds even on a phone.

Document mode does that to each color separately, which is the reason a blue pen stays blue and a red stamp stays red while the paper around them goes white. Grayscale does the same and drops the color. Black and white does it and then decides one threshold for the whole page, for the times somebody wants the smallest, cleanest possible copy. And the photograph as it is does none of it, for the times the picture is evidence and the lighting is part of what it shows.

What it will not do, and the big one is text

The PDF this makes holds pictures of your pages. The words in it are not text and cannot be searched, selected or copied, and nothing about the way the page looks will tell anybody that until they try. This site will not run character recognition in a browser and pretend otherwise, because the honest version of that needs a real engine and a real dictionary and it would be wrong often enough to be worse than useless on a handwritten date.

So a scan is the right thing for paper that was signed by hand and has to be filed, and it is the wrong thing for a document that already exists as a file. If you are photographing a page you printed off your own computer, go back and print it to PDF instead. The file will be a tenth the size, the words will be words, and it will look better.

It cannot recover what the camera did not see. If the sheet runs off the side of the picture, the corner it hands you is the corner of the photograph and not the corner of the paper, and the page comes out with an edge of the document missing. That case is caught and said out loud rather than quietly delivered, because the result looks fine on screen and the mistake only shows up later, in the packet, at the worst possible moment.

And it does not fix a bad photograph. A page that is out of focus stays out of focus, a page photographed in the dark stays noisy, and a page with a coffee cup sitting on it comes out with the coffee cup. Nothing here invents detail that was not in the picture. Take it again with the whole sheet in frame, in the open shade rather than in direct sun, and the second attempt takes ten seconds.

Where the file size goes

A scanned page is a picture, and a picture is heavier than text, so a twelve page scan is a bigger file than a twelve page document. What it is not is a twelve megapixel photograph twelve times over. Each page is held at a size that suits paper rather than a size that suits a camera sensor, which is where nearly all of the weight of a phone photo goes, and the result is usually a few hundred kilobytes a page rather than several megabytes.

Black and white is by far the smallest, because a page that is only ever black or white compresses into almost nothing. It is also the mode most likely to lose a light pencil note or a faint carbon copy, so it is worth looking at the preview before choosing it on anything that has to be read rather than merely filed.

If the finished file is still too heavy for somebody's mail server, make it smaller. That re-encodes the pictures inside it, which on a scan is where every byte is.

Getting a good photograph in the first place

Light beats everything else. Open shade is the best light there is for this: under the tailgate, inside the doorway, on the shaded side of the truck. Direct sun is the worst, because a sheet of white paper in full sun is brighter than the camera can hold and the writing goes flat and gray. Indoors, turn so the window is in front of you rather than behind you, so your own shadow falls behind you and not on the page.

Fill the frame with the sheet and leave a couple of inches of whatever it is lying on visible all the way round. The edges of the paper are how the sheet is found, so a page that runs out of the picture is a page whose corners were never in it. And put it on something that is not the same color as paper. A white sheet on a white table is the one case that genuinely defeats this, and moving it onto the truck bed or the ground fixes it in two seconds.

Hold the phone flat over the page rather than leaning in over it. The angle is corrected either way, but a shot taken from directly above needs less correcting and stays sharper across the whole sheet. And tap the screen on the middle of the page before you take it, so the camera focuses on the writing rather than on the wood grain next to it.

Then take them all in one go. The pages come in as a set, in the order you picked them, and any one of them can be taken back out from the strip of thumbnails without starting again.

What happens around it

This is usually the first step and not the last one, because it is how paper gets into a file at all. What comes next is whatever you would do with any other document.

Look at the edges before it goes anywhere. A photograph taken on a desk catches what else was on the desk, and a page photographed on a clipboard catches the page underneath it. Nobody examines the border of their own scan. If something is in there that should not travel, cut it off or, if it is somebody's pricing or an account number, black it out, which is the only one of the two that actually removes it.

Then put it in the packet with the invoice and the photographs, number the pages so a question about it can be pointed at by page, and send it.

And if what you are scanning is a waiver, the better move is usually not to scan it at all. Fill out the statutory form here, send it as a file, and get it back signed as a file. A scan is the fallback for paper that already exists, not the way paper should start.

Questions people ask

Related