How to remove a background from a photo, a signature, or a logo

Two different problems wearing one name

Cutting a person out of a photo and cutting a signature out of a scanned page are not the same task, even though both get called "background removal." A photo needs something that can recognize a person against a complex, irregular background — hair, clothing edges, uneven lighting. A signature needs almost the opposite: it's already high-contrast ink on a flat background, and the "hard" part is just telling ink from paper accurately at the edges. Using a person-detection tool on a signature doesn't produce a bad cutout — it correctly finds no person at all and gives you back nothing useful, because that's genuinely what it was asked to look for.

Removing a background from a photo of a person

This needs a model trained to recognize people — technically called semantic segmentation, but practically it just means the software has learned what a person silhouette looks like against thousands of different backgrounds and lighting conditions. It works well on portraits, selfies, and headshots specifically because that's what it was trained on. It works poorly or not at all on a product photo, an animal, or a landscape, because none of those are "a person" as far as the model is concerned — that's a real, structural limitation, not a bug to be tuned away.

The most common quality issue with photo cutouts is soft or fuzzy hair edges. That's usually intentional in a well-built tool: a hard, pixel-perfect edge around individual hair strands looks more artificial than a slightly soft blend, the same way a feathered edge looks more natural than a razor-sharp cutout in traditional photo editing.

Removing a background from a signature, logo, or scanned document

This is a completely different, and honestly simpler, problem: detect the background color and make everything close to it transparent, while keeping everything far from it (the ink, the graphic) opaque. No model, no training, no "recognizing" anything — just comparing each pixel's color distance from the background. It's also near-instant, since there's no model to load or run.

The one setting worth understanding here is sensitivity — how far a pixel's color has to be from the detected background before it counts as "subject, not background." Too low, and faint parts of a light signature or a soft shadow around a logo stay visible. Too high, and it starts eating into the edges of the ink itself, especially on strokes that trail off thin. A quick way to dial it in: raise sensitivity until background artifacts disappear, then back off slightly the moment you see the edges of the actual subject start to erode.

Getting a clean scan before you even start

For signatures and logos specifically, the input quality matters more than any setting downstream. A signature photographed at an angle, under uneven lighting, with a shadow falling across half the page, gives the background-removal step a background that isn't actually one uniform color — which is exactly the assumption it relies on. Flat, even lighting and a straight-on scan or photo will get a noticeably cleaner cutout than any amount of sensitivity tweaking can recover from a poor original.

What "transparent background" actually means for the file

A transparent background only survives in formats that support an alpha (transparency) channel — PNG and WebP, not JPEG. If you need the cutout to actually show as transparent later (dropped onto a letterhead, layered in a design tool, used as a stamp), make sure whatever you download or convert it to afterward stays in PNG, since converting a transparent PNG to JPEG fills the transparent area with solid white instead.