Most OCR failures don't come from the model — they come from the photo. Before you reach for an "enhance" button or a sharpening filter, fix the input. The single highest-leverage habit is boringly simple: flatten the page and shoot it square.
Every degree of skew and every curl in the paper forces the OCR engine to guess at character shapes that are no longer where it expects them. A model that reads a flat, well-lit page at 99% will drop several points on the exact same text photographed at an angle on a wrinkled surface.
Why flat and square wins
Modern OCR is remarkably tolerant of bad lighting and low resolution, but it is far less forgiving of geometric distortion. When a line of text bows or tilts, the spacing between characters stops being uniform, and letters that rely on subtle differences — like rn versus m, or 5 versus S — start to blur into each other from the engine's point of view.
Getting the geometry right removes that whole class of errors up front:
- Lay the document on a hard, flat surface. A table beats your lap; a clipboard beats a stack of loose paper.
- Hold the camera parallel to the page, not tilted over it. Look for the page edges to run straight and level in the frame.
- Fill the frame with the document so the text uses as many pixels as possible, but leave a small margin so nothing is clipped.
- Kill glare by shooting in soft, even light instead of under a single harsh lamp.
Do this, then let the tool do the rest
Once the page is flat, square, and evenly lit, you can stop fussing. Upload it, let the scanner extract the text, and spend your attention on verifying the handful of fields that actually matter — totals, names, dates — rather than fighting a distorted image.
Want to see the difference on your own worst document? Open the scanner and run one photo the old way, then reshoot it flat and square. The cleaner input almost always speaks for itself.