Phase 2: ROI candidate detector for code-like regions in beams

[ProhibitedTV]

Parent roadmap: #3

Objective

Detect regions of interest before OCR so the app can focus on plausible code-like structures instead of running OCR blindly across full noisy frames.

Tasks

• Add ROI detectors for connected components, line/edge density, repeated glyph-like contours, high-frequency texture bands, and grid-like arrangements.
• Score ROIs by structure density, persistence across nearby frames, and contrast against local background.
• Store ROI bounding boxes, masks, crops, detector scores, and parent preprocessing variant.
• Add non-max suppression / deduplication for overlapping ROIs.
• Render ROI overlays in the dashboard and HTML report.

Acceptance criteria

• Every OCR attempt references a specific ROI or explicitly states full_frame.
• Top candidates include raw frame, processed frame, ROI crop, and overlay.
• ROI detector behavior is tested on synthetic positives, synthetic negatives, and real fixture media.

Notes

This is the bridge between optical analysis and OCR. It should reduce hallucinated OCR by making the candidate region explicit.