gbanyan
723a3f6eaf
Major §III restructuring after codex rounds 29-34 demolished the distributional path to thresholds (Scripts 39b-39e prove (cos, dHash) multimodality is composition-driven + integer-tie artefact). v4.0 pivots to anchor-based multi-level inter-CPA coincidence-rate (ICCR) calibration via Scripts 40b, 43, 44, 45, 46: - §III-G: scope justification rewritten (LOOO + Firm A case study + within-firm collision structure; dropped "smallest scope rejects unimodality" rationale); added sample-size reconciliation (150,442 descriptor-complete vs 150,453 vector-complete; 437 accountant-level vs 468 all) - §III-I: new sub-section I.4 composition decomposition (2x2 factorial centred + jittered Big-4 pooled dh p=0.35); I.5 conclusion of no natural threshold - §III-J: K=3 recast as firm-compositional descriptive partition (not three mechanism clusters); bridge to §III-L.4 cross-firm hit matrix added - §III-K: Score 1 reframed as firm-composition position score - §III-L: NEW major sub-section — anchor-based threshold calibration with L.0 methodology, L.1 per-comparison ICCR (replicates v3 cos>0.95 -> 0.0006; new dh<=5 -> 0.0013; joint -> 0.00014), L.2 pool-normalised per-signature ICCR (any-pair HC 11.02%; per-firm A 25.94% vs B/C/D <1.5%), L.3 doc-level ICCR (HC 18%; HC+MC 34%), L.4 firm heterogeneity logistic OR 0.01-0.05 + cross-firm hit matrix (98-100% within-firm), L.5 alert-rate sensitivity (HC threshold locally sensitive not plateau-stable), L.6 observed deployed alert rate excess over inter-CPA proxy - §III-M: NEW sub-section — multi-tool validation strategy under unsupervised setting; 9 partial-evidence diagnostics each with disclosed untested assumption; positioning as anchor-calibrated screening framework with human-in-the-loop review, NOT validated forensic detector - Terminology: "FAR" replaced with "inter-CPA coincidence rate (ICCR)" throughout; primary metric name change documented in §III-L.0 - Provenance table: ~35 new rows for Scripts 39b-e/40b/43-46; "key numerical claims" instead of "every numerical claim" - Removed v2-v6 internal changelog metadata; v7 draft note added Codex round-32 SOUND_WITH_QUALIFICATIONS, round-33 GO_WITH_REVISIONS, round-34 READY_WITH_NARROW_FIXES (all 8 patches applied). Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
PDF Signature Extraction System
Automated extraction of handwritten Chinese signatures from PDF documents using hybrid VLM + Computer Vision approach.
Quick Start
Step 1: Extract Pages from CSV
cd /Volumes/NV2/pdf_recognize
source venv/bin/activate
python extract_pages_from_csv.py
Step 2: Extract Signatures
python extract_signatures_hybrid.py
Documentation
- PROJECT_DOCUMENTATION.md - Complete project history, all approaches tested, detailed results
- README_page_extraction.md - Page extraction documentation
- README_hybrid_extraction.md - Hybrid signature extraction documentation
Current Performance
Test Dataset: 5 PDF pages
- Signatures expected: 10
- Signatures found: 7
- Precision: 100% (no false positives)
- Recall: 70%
Key Features
✅ Hybrid Approach: VLM name extraction + CV detection + VLM verification
✅ Name-Based: Signatures saved as signature_周寶蓮.png
✅ No False Positives: Name-specific verification filters out dates, text, stamps
✅ Duplicate Prevention: Only one signature per person
✅ Handles Both: PDFs with/without text layer
File Structure
extract_pages_from_csv.py # Step 1: Extract pages
extract_signatures_hybrid.py # Step 2: Extract signatures (CURRENT)
README.md # This file
PROJECT_DOCUMENTATION.md # Complete documentation
README_page_extraction.md # Page extraction guide
README_hybrid_extraction.md # Signature extraction guide
Requirements
- Python 3.9+
- PyMuPDF, OpenCV, NumPy, Requests
- Ollama with qwen2.5vl:32b model
- Ollama instance: http://192.168.30.36:11434
Data
- Input:
/Volumes/NV2/PDF-Processing/master_signatures.csv(86,073 rows) - PDFs:
/Volumes/NV2/PDF-Processing/total-pdf/batch_*/ - Output:
/Volumes/NV2/PDF-Processing/signature-image-output/
Status
✅ Page extraction: Tested with 100 files, working ✅ Signature extraction: Tested with 5 files, 70% recall, 100% precision ⏳ Large-scale testing: Pending ⏳ Full dataset (86K files): Pending
See PROJECT_DOCUMENTATION.md for complete details.