Document Scanner

Document scanning used to mean a flatbed the size of a small printer, plugged into a desktop, taking thirty seconds per page. Phones killed that workflow somewhere around 2015 — modern cameras shoot 12+ megapixels with sensors good enough that the limiting factor for a paper scan is no longer image quality but the geometric and tonal post-processing that turns a snapshot of a page into something that looks scanned. Apps like CamScanner and Adobe Scan do that processing well, but the cost is obvious: accounts, paywalls, cloud archives tied to your identity, and a whole app to install for a one-shot operation. This document scanner does the same job from a single web page — no account, no install — and your photo lives only in memory on our server for the few seconds it takes to process.

The pipeline has four stages: capture, edge detection, perspective correction, and tonal enhancement. Capture is straightforward — point the camera at the document, frame it loosely (the corners do not need to be exactly aligned), tap the shutter. Edge detection finds the four corners of the page by looking for the strongest contrast boundaries that form a quadrilateral; when it is uncertain (busy backgrounds, white-on-white surfaces) you can drag the four corner handles manually. Manual override is fast and exact, and it is the right answer when auto-detect lacks confidence — the rest of the pipeline trusts whichever corners you give it.

Perspective correction is the step that makes the difference between a photo of a page and an actual scan. When you photograph a document on a desk from above, the resulting image has a trapezoidal distortion: the side closer to the camera looks larger than the side further away, even though both sides of the page are the same width in reality. Mathematically, this is a projective transformation between the document plane and the camera sensor plane. Reversing it requires computing a homography — a 3x3 matrix that maps the four detected corners back to a clean rectangle. The output pixel grid is then sampled bicubically through the matrix. This step runs on our server using OpenCV and takes a few hundred milliseconds for a typical full-resolution phone photo.

Tonal enhancement is the last stage and the one that distinguishes a usable scan from a phone photo. Four filter modes cover real scanning. Color keeps the colors but corrects perspective; Magic Color (the default) removes shadows, whitens the paper, lifts contrast and sharpens text — this is the office-scan look. Grayscale converts to a perception-weighted gray, the right choice for printed text where color carries no information. B&W uses an adaptive threshold that walks each pixel and compares it to its 31x31 neighborhood mean; this beats a global threshold because it handles documents with uneven lighting and produces the crisp pure-black-on-pure-white look most professional scanners default to for forms and contracts.

This release is intentionally single-page focused: capture one sheet, optionally drag the corners, scan, download. When you need a multi-page document, scan each page separately and join them with [Images to PDF](/tools/images-to-pdf).

Edge cases worth knowing. Glossy paper under direct overhead light produces glare that no filter can remove because the underlying pixels are saturated white; reposition the page or your light source. Lined notebook paper with strong horizontal lines can confuse auto edge detection; drag the corners manually in that case. Pages with photographs inside can over-threshold in B&W; use Grayscale or Magic Color instead. Curled or folded documents will keep some distortion in the middle because a single photo cannot recover the 3D shape; flatten the page under a book before scanning if absolute fidelity matters.

Mobile considerations are non-negotiable here because most users will reach for their phone first. The interface uses native pointer events so finger drags work identically to mouse drags, the corner handles are sized for touch, and the bottom action bar is fixed so primary actions stay reachable with the thumb. Camera capture defaults to the back camera (`facingMode: environment`). After capture, the photo immediately enters the corner-adjust view with auto-detect on — most documents need zero manual adjustment, but the corners are draggable for the cases that do.

Output is a standards-compliant PNG that opens in Adobe Acrobat, Apple Preview, Google Drive, your file manager and every image viewer. If you prefer pure black/white or grayscale, switch the filter chip before scanning; Magic gives you the office-scan look with sharpened text and a clean white background.

Privacy: capture and corner selection run in your browser. Cropping, perspective correction, deskew and paper whitening run on our server inside an in-memory OpenCV pipeline. Your image is not written to disk and not logged; it lives only for the few hundred milliseconds the request takes, then it is dropped. No account, no archive, no per-user tracking. Open the network tab during a scan and you will see one request for the scan endpoint and zero retained state on our end after the response.

The tool pairs naturally with the rest of Omnvert's PDF and image utilities. Combine multiple scans into one document with [Images to PDF](/tools/images-to-pdf), shrink the file with [PDF Compress](/tools/pdf-compress), add a searchable text layer with [PDF OCR](/tools/pdf-ocr), or sign with [PDF Sign](/tools/pdf-sign). For a different intake — say, a flatbed scan already in PDF form — the [PDF OCR](/tools/pdf-ocr) tool can ingest that directly.

A few practical tips. Diffuse overhead light produces the cleanest results — direct sunlight or a single-source desk lamp creates harsh shadows that survive even tonal correction. Place the page on a surface that contrasts with the paper color so auto-detect has an easy time (white paper on a wood desk works well; white paper on a white tablecloth fights the algorithm). Leave a small margin around the document in the frame so the corners are clearly visible and not cropped. Hold the phone parallel to the page if possible — perspective correction handles modest angles fine, but extreme angles (45° tilt) lose resolution because the far side of the page is much smaller in the source pixels than the near side.

Comparison with traditional flatbed scanners. A flatbed produces marginally better results for absolute fidelity (uniform lighting, perfect flatness, calibrated color), but the workflow gap is enormous: a phone scan takes 5–8 seconds per page; a flatbed takes 20–30 seconds per page from lid-open to scanned image. For the vast majority of everyday scanning — receipts, contracts, ID copies, recipe pages, school handouts — a phone scan is good enough and several times faster. Flatbeds remain the right tool for irreplaceable documents where color accuracy matters.