Boundary, Wand, and RGB Deviation¶
Status: RGB-deviation control accepted for UI; final D and semantic island quality not accepted
Authority: Accepted decision DL-010 updates from 2026-08-01
Last reviewed: 2026-08-14
Question¶
What did boundary detection, magic-wand methods, and RGB deviation teach the lab?
Short answer¶
Boundary evidence can exclude a title block when a long divider looks like a frame. Magic-wand experiments showed desktop-tool semantics matter. The owner accepted one direct RGB-deviation control for the UI. That acceptance does not validate final extraction quality.
Boundary detection¶
What we tried¶
The unchanged detector retained lines spanning at least 65% of the page.
It searched the outer 25% side bands.
It chose the innermost candidate and moved two pixels inward.
Why it helped¶
It created a stable scan rectangle for cleanup experiments.
Why it failed or stopped¶
A long title-block divider can satisfy the rule and exclude the title block. The working area is now an acceptance concern, not a trusted preprocessing fact. Detector change remains stopped pending representative owner review and a fixed reference-independent rule.
What we retained¶
Visible boundary-evidence diagnostics and the requirement that content outside the working rectangle must not silently disappear.
Magic-wand methods¶
What we tried¶
Edge-connected white removal. Largest-flow arms. Pinta-style and Inkscape-style clicked-colour comparisons.
Why it helped¶
Pinta-style was closest on the two supplied owner masks.
Reported in-scope IoU values were 0.999981462 and 0.947326804 with zero in-scope over-removal.
Why it failed or stopped¶
Only two wand masks were validated. Semantic island accuracy remained unknown. App-style wand labels were replaced by a direct control. Phase 4C1D stayed stopped.
What we retained¶
Flood-semantics lessons and history of older methods for review.
RGB deviation¶
What we tried¶
One explicit control with white fixed at RGB(255,255,255).
Deviation D in 0..30.
Every channel must be at least 255-D.
Selected removal is the four-connected eligible component from one fixed automatic or manual click.
Why it helped¶
It replaced opaque Pinta or Inkscape choices with one reviewable control. Determinism and control checks passed.
Why it stopped short of acceptance as extraction truth¶
The final D still requires owner selection. Semantic quality of resulting islands is not validated. Phase 4C1D remains stopped.
What we retained¶
The accepted UI and control design for direct paper-color deviation.
Open Questions¶
- Which D value will the owner accept for default Extraction cleanup?
Sources¶
- Accepted decision: DL-010 (2026-08-01 RGB, wand, and working-area updates)
- Phase 4C1C evaluation reports cited by DL-010