Image Extraction Skill

Extract illustrations from historical book scans using Gemini vision AI. Creates a curated gallery with rich metadata for each image.

Invocation

/extract-images                      # Show extraction status across library
/extract-images BOOK_ID              # Extract images from specific book
/extract-images --all                # Extract from all books missing images
/extract-images --priority           # Extract from high-priority books first

What Gets Extracted

For each illustration found:

• Bounding box - Precise crop coordinates
• Type classification - emblem, woodcut, engraving, portrait, frontispiece, musical_score, diagram, symbol, decorative, map
• Gallery quality score - 0.0-1.0 rating for curation
• Museum description - 2-3 sentence scholarly label
• Structured metadata:
  • subjects - Topics (alchemy, transformation, mythology)
  • figures - People/beings depicted (Mercury, serpent, old man)
  • symbols - Symbolic elements (ouroboros, athanor, philosophical egg)
  • style - Art historical style (Northern European Renaissance)
  • technique - Production method (woodcut, engraving)

API Endpoints

Check Extraction Status

# Get books with extraction counts
curl -s "https://sourcelibrary.org/api/books" | jq '[.[] | {
  id,
  title: .title[0:40],
  pages: .pages_count,
  images: (.detected_images_count // 0)
}] | sort_by(-.pages)'

# Find books needing extraction (no detected_images)
curl -s "https://sourcelibrary.org/api/books" | jq '[.[] |
  select((.detected_images_count // 0) == 0) |
  {id, title: .title[0:40], pages: .pages_count}
] | sort_by(-.pages)'

Extract Images from a Book

Smart Extraction (Recommended)

Uses OCR <image-desc> tags to find pages with images, ranks them with AI, and only extracts the best 15-20:

# Extract top 20 images from a book
node scripts/smart-extract.mjs BOOK_ID --limit=20

# Extract top 15 images
node scripts/smart-extract.mjs BOOK_ID --limit=15

How it works:

1. Finds pages with <image-desc> tags in OCR (from previous OCR pass)
2. Uses AI to rank descriptions and pick the most interesting
3. Only extracts those specific pages
4. ~30 seconds for 20 images instead of scanning all pages

Full Extraction (All Pages)

For comprehensive extraction of every page (slow, use sparingly):

node scripts/evaluate-extraction.mjs BOOK_ID --clear

Via API (Individual Pages)

curl -s -X POST "https://sourcelibrary.org/api/extract-images" \
  -H "Content-Type: application/json" \
  -d '{
    "bookId": "BOOK_ID",
    "pageIds": ["PAGE_ID_1", "PAGE_ID_2"],
    "overwrite": false
  }'

Books with OCR Image Descriptions

Check which books have <image-desc> tags (best for smart extraction):

# Find books with image descriptions in OCR
node -e "
const { MongoClient } = require('mongodb');
require('dotenv').config({ path: '.env.local' });
// ... aggregation to find books with <image-desc> tags
"

Output Quality

Gallery Quality Scoring (0.0-1.0)

Key Rule: Images with people/figures always score 0.8+

Image Types

Viewing Results

Gallery Page

https://sourcelibrary.org/gallery?bookId=BOOK_ID

Book Guide Page (high-quality images only)

https://sourcelibrary.org/book/BOOK_ID/guide

Individual Image Detail

https://sourcelibrary.org/gallery/image/PAGE_ID:INDEX

Cost Estimation

Using Gemini 2.5 Flash:

Monitoring Extraction

# Check extraction results for a book
curl -s "https://sourcelibrary.org/api/gallery?bookId=BOOK_ID" | jq '{
  total: .total,
  high_quality: [.images[] | select(.galleryQuality >= 0.75)] | length,
  types: [.images[].type] | group_by(.) | map({type: .[0], count: length})
}'

# Sample image metadata
curl -s "https://sourcelibrary.org/api/gallery/image/PAGE_ID:0" | jq '{
  type,
  galleryQuality,
  museumDescription,
  metadata
}'

Troubleshooting

Rate Limiting

The evaluation script has built-in 500ms delays. For higher throughput:

• Use multiple API keys (rotate on 429)
• Reduce delay for paid tiers

Missing Images

If extraction finds fewer images than expected:

• Check page image quality (blurry scans)
• Verify pages have cropped_photo or photo_original URLs
• Re-run with --clear to retry

Wrong Classifications

If images are misclassified:

• Edit via gallery UI at /gallery/image/PAGE_ID:INDEX
• Use PATCH endpoint to update type/metadata