Everything Talpatra Scan's offline document scanner does
Six capabilities, all running on the phone in your hand. No step in this pipeline calls out to a server, because the app is built to be unable to.
Boundary detection and auto-capture
Talpatra Scan finds the page edges in the live camera feed and fires the shutter itself, so you don't have to line up a rectangle by hand or guess when the shot is steady.
Two detectors run in a primary/fallback pair on every frame: a classical computer-vision contour search, and a corner-heatmap ML model that kicks in when the classical pass loses the edges — low contrast, a patterned desk, a curled corner. A stabilizer watches the detected boundary across frames and only triggers capture once it holds steady, so you get a clean shot instead of a blurry guess.
Book and curved-page dewarp
Photograph an open book or a page that won't lie flat, and Talpatra Scan straightens the curve back out — entirely offline. This is the one capability most free, ad-supported scanners skip, because it's expensive to build and doesn't run well without a network round trip to do the heavy lifting.
Correction runs as a composite of two engines: a classical geometry pass that models the page's curvature from detected text lines, and a neural dewarp model (built on UVDoc, a published mesh-grid unwarping architecture) that predicts a correction grid directly from the image. The two are combined rather than swapped — the geometric model is fast and predictable, the neural model catches distortion the geometry alone misses. Both run on the device's CPU; nothing is uploaded to make this work.
On-device OCR, searchable PDFs
Every scan gets a real, selectable, searchable text layer — recognized on the phone, never sent anywhere to be read.
Talpatra Scan runs Google's ML Kit text recognizer against the corrected page image and positions each recognized word at its actual location on the page, so the exported PDF's invisible text layer lines up with what you see — not just a blob of text dumped at the top. Search, copy, and paste all work the way they would on a PDF exported from a desktop scanner.
Filters and export profiles
Different scans need different output: a novel page reads best clean and light, a form needs to stay legible after compression, a receipt just needs to be small. Talpatra Scan lets you pick the render per document instead of forcing one look on everything.
Grayscale is the default — full tonal range, small file size, safe for photos and mixed content. A black-and-white "document" mode flattens uneven lighting and thresholds the page to pure black text on white, built for text-only pages headed somewhere that only cares about crisp glyphs. A "reading" mode applies that same lighting flattening without the hard threshold, for a page that looks clean on a screen but keeps its grayscale softness.
Duplicate-page detection
Scan the same page twice by accident — easy to do mid-stack — and Talpatra Scan flags it before it ends up buried in a 40-page export.
Each page gets a perceptual hash (dHash) computed from its image content. Pages with near-identical hashes are flagged as likely duplicates so you can review and remove the redundant shot before you export, instead of finding it later in the finished PDF.
Non-destructive editing
Crop too tight or pick the wrong filter, and you can always get back to the original — every edit in Talpatra Scan is reversible, because the app never throws away the source photo.
The original capture and the exact crop boundary used to process it are kept alongside every page, not overwritten by later edits. Reopen a page, adjust the crop or switch filters, and it reprocesses from that original — it doesn't compound corrections on top of an already-edited image.
All of it runs without the internet permission
Not a setting you have to turn off — Talpatra Scan is built withoutandroid.permission.INTERNET, enforced by a build step that fails if it ever reappears.