Purpose
Restrict recognition to the information the user is trying to read.
Role in the complete pipeline
Detection removes irrelevant surroundings, estimates page geometry, and provides framing evidence to the guidance loop.
INPUT
→A camera frame and its quality measurements.
METHOD
- Find edges and candidate contours
- Approximate page corners
- Correct perspective where possible
- Locate the primary connected text region
A page boundary, a normalized crop, and detection confidence signals.
Common failure cases
- Page and background have similar contrast
- Only part of the page is visible
- Curved paper breaks the rectangular assumption
- Busy backgrounds create false regions
How VisoraAI responds
- Fall back to text-region localization
- Ask the user to center the page
- Reject unstable crops instead of forwarding them
Current limitations
- Contour methods are sensitive to lighting
- Curved and folded documents remain difficult
Planned improvements
- Evaluate learned document detectors
- Improve partial-page handling
- Track corners across consecutive frames