Academic figure architecture extractor

Academic Figure Architecture Extractor & Analyzer

Turn paper PDFs or existing architecture images into a structured 架构图分析结果 that handoffs cleanly to color-expert and prompt skills.

Palettes: → ../docs/palettes.md (names only)
Missing info: → ../docs/missing-info-policy.md
Extractor: → scripts/extract_pdf_figures.py

Honest scope

• Prefer user-supplied figure images when available.
• PDF extraction is a real local helper (pdfimages / PyMuPDF / optional pdftoppm), not a trained detector.
• Size filter is heuristic only; architecture vs photo/table is agent judgment.
• Palette: recommend names from docs/palettes.md; hex via color-expert when needed.

Input Contract

• Prefer: PDF path(s), figure images, domain, venue
• Minimum: one PDF or one architecture image
• Missing: analyze what exists; list blocked steps

Output Contract — 架构图分析结果

• inventory (path/page, size, keep/drop reason)
• per-kept-figure structure (components, hierarchy, flow, type)
• recommended palette names
• redraw parameters for academic-figure-prompt

Steps

Step 1: Obtain images

Images given → index paths.

PDF given → run the helper (from skill dir or repo root):

python3 academic-figure-architecture-extractor/scripts/extract_pdf_figures.py \
  /path/to/paper.pdf -o /tmp/arch-extract/paper

Useful flags:

flag	meaning
--backend auto|pdfimages|pymupdf	embedded-image backend
--min-side 300	drop tiny icons (default)
--min-pixels 90000	drop low-res crops
--pages 4 or --pages 1-3 or --pages all	also rasterize pages via pdftoppm
--dpi 150	raster DPI

Read extract-report.json in the out dir (kept / dropped / tools).

If both backends missing → ask user for exported figures; do not invent paths.

Done when: each candidate has a path or page reference, or a clear tool blocker is stated.

Step 2: Filter to architecture-like figures

Start from kept (size-ok). Agent reclassifies:

keep cues	drop cues
boxes + arrows, layered blocks	pure photos, scatter-only, dense tables
structured edges / modules	tiny icons already size-dropped

Unsure → keep + 待确认.

Done when: each image is keep / drop / uncertain with a one-line reason.

Step 3: Structure analysis

For each kept figure:

1. components (core vs auxiliary)
2. hierarchy / dataflow
3. type: Overall Framework / Network Architecture / Module Detail / Comparison
4. domain notes

Done when: every kept figure has type + component list + flow summary.

Step 4: Palette suggestion + redraw handoff

Map via docs/palettes.md (e.g. ≥4-module framework → Nature Blue; module detail → Blue Monochrome; comparison → ML TopConf Deep).

图类型: ...
核心组件: ...
配色方案名: ...
布局建议: 16:9 | 3:2 | 4:3
风格: white fill, colored borders, flat vector
标注要求: ...

Done when: each kept figure has redraw params + palette names (no hex tables).

Stop

Stop when the report is delivered, extraction is blocked pending user images, or the user only wanted inventory.