AI Auto Cropper DocsDetection
How photos are found, the detection methods, and tilt handling.
Detection works on a downscaled copy of each scan (about 1200 px on the long side). Boxes are then mapped back to the original pixels, so exports are always cut from the full-resolution file.
How it finds photos
The scan's border is sampled to learn the background colour. Pixels that differ from it — or that sit on a strong edge — are grouped into regions. Regions that look like prints (a sensible size, shape and fill) become boxes. For each box a rotated rectangle is fitted to the print's own pixels, so a print lying at an angle gets a tilted box, and the box is pushed out to the print's real edge. If no separate print is found, the scan is treated as an ordinary photo and handed to AI subject detection (below). If that finds no clear subject either, the whole image is kept as one photo.
Methods (Advanced options)
Choose the method before adding scans. The choice is saved in this browser and applies to scans you add next.
| Method | Use it for |
|---|---|
| Auto (default) | Colour + edge detection. Best for prints on a scanner bed or an album page. |
| AI subject | Ordinary photos (a pet on the floor, a portrait). Always crops around the main subject. |
| White / plain background | Colour only. Clean light or dark bed with gaps between prints. |
| Flexible background | Edge driven, tolerates patterned pages (scrapbooks, magazines). May add extra boxes to delete. |
| Single large image | One photo almost filling the scan. At most one box. |
| Transparent PNG | Boxes the non-transparent part of a cut-out PNG. Always one box. |
AI subject detection
For ordinary photos — not several prints — an object-detection model (EfficientDet-Lite0) and a segmentation model (DeepLab v3) run in your browser via MediaPipe. They download once (about 10 MB, plus the 13 MB runtime) the first time they are needed; your photo is never uploaded.
- The main subject is the most confident, largest, most central detection; people and animals are preferred over furniture.
- The box is fitted to the subject's outline with padding of about 12%. When the subject is clearly elongated and lying at an angle (a dog stretched out diagonally), the box is tilted to match, and export straightens it.
- The segmentation knows people, cats, dogs, birds, horses, sheep, cows and some vehicles and furniture. Other objects (about 80 everyday kinds) get a straight box from the detector.
- It is a suggestion: if no subject is found, or the crop would keep almost the whole frame, the whole image is kept. Adjust or delete the box like any other. A crop made this way is labelled AI subject in the crop list.
In Auto, this runs only when the print detector finds a single full-frame photo, so multi-print scans are unaffected.
Tilt and orientation
- Straightening: each box carries the tilt of its print. Export rotates the crop so the photo comes out level.
- Ignore minor rotations (on by default): tilts under 2° are treated as level, because tiny rotations soften an image slightly.
- Sideways or upside-down photos are not detected automatically. Use Rotate left / right on a crop to turn it in 90° steps.
Tips for good detection
Use a plain, even background, leave a finger-width gap between prints, and keep glossy sleeves from reflecting. 300–600 DPI gives crops you can reprint.