Field guide
How a text detector finds words in pictures
Text detection and text recognition are separate jobs. A region can look like lettering even when the model cannot reliably name the characters inside it.
01
Detectors search for aligned strokes, contrast and word-like geometry across multiple scales. Fences, windows and patterned fabric can produce the same cues.
02
Recognition confidence helps reject false regions but can also miss stylised, curved or low-contrast lettering. Inspect the proposed mask rather than trusting a threshold alone.
03
Removal requires a mask wider than the visible dark stroke because outlines, shadows and antialiasing occupy neighbouring pixels. Too wide a mask consumes genuine picture detail.
04
Keep detection and fill reversible until review. A plausible patch can hide the fact that the mask crossed a face, product edge or legally important label.