OCR PNG

PNG is the best-case input. It is lossless, so every pixel that was captured is still there, with no compression ringing around character edges and no blocky artefacts to smear one letter into its neighbour. That matters because it is also the default screenshot format on most systems, which means PNG input is usually machine-rendered text: perfectly aligned, uniformly lit, anti-aliased by the operating system rather than blurred by optics. Recognition on that kind of image approaches its ceiling — the remaining errors tend to be genuine ambiguities like a capital I against a lowercase l in a sans-serif typeface, rather than image quality problems. PNG also supports transparency, which needs handling: a transparent background is composited onto white before recognition, since otherwise dark text on an empty alpha channel can be read as dark on dark. Light text on a dark background is inverted first, as recognition assumes dark ink on light paper.