Extract PDF Figure

PDF Figure & Table Extraction Skill

Extract figures and tables from academic PDF papers with pixel-perfect precision using the Qwen3-VL vision-language model. Features multi-round quality assessment for optimal extraction accuracy.

Quick Start

# Extract a complete figure (including title, legend, notes)
python .claude/skills/extract-pdf-figure/scripts/extract_figures.py "<pdf_path>" "<figure_name>"

# Extract main content only (no title, legend, notes)
python .claude/skills/extract-pdf-figure/scripts/extract_figures.py "<pdf_path>" "<figure_name>" --no-extras

# Batch extraction with output directory
python .claude/skills/extract-pdf-figure/scripts/extract_figures.py "<pdf_path>" --batch "Figure 1,Figure 2,Table 1" -d "output/asset/"

Features

1. Extraction Modes

2. Multi-Round Quality Assessment

The tool uses an iterative refinement process:

1. Round 1: Initial figure detection and bounding box estimation
2. Round 2+: Quality assessment with visual feedback (red bounding box overlay)
3. Refinement: If quality score < 8/10, coordinates are automatically adjusted
4. Termination: Stops when quality is satisfactory or max rounds reached (default: 3)

3. Supported Figure Types

• Figure 1, Figure 2, etc.
• Table 1, Table 2, etc.
• Fig. 1, Fig 1 (abbreviated)
• Figure 1(a), Figure 1 (a), Figure 1a (sub-figures)
• Table 1(a) (sub-tables)

Usage Examples

Single Figure Extraction

# Complete figure with all elements
python .claude/skills/extract-pdf-figure/scripts/extract_figures.py "PDFs/paper.pdf" "Figure 1"

# Only the chart content (no caption)
python .claude/skills/extract-pdf-figure/scripts/extract_figures.py "PDFs/paper.pdf" "Figure 1" --no-extras

# With custom output path
python .claude/skills/extract-pdf-figure/scripts/extract_figures.py "PDFs/paper.pdf" "Table 2" -o "output/asset/Table_2.png"

# Higher resolution
python .claude/skills/extract-pdf-figure/scripts/extract_figures.py "PDFs/paper.pdf" "Figure 3" --dpi 400

Sub-Figure Extraction

# Extract only Figure 1(a) from a composite figure
python .claude/skills/extract-pdf-figure/scripts/extract_figures.py "PDFs/paper.pdf" "Figure 1(a)"

# Sub-figure with custom output
python .claude/skills/extract-pdf-figure/scripts/extract_figures.py "PDFs/paper.pdf" "Figure 2(b)" -o "output/asset/Figure_2b.png"

Batch Extraction

# Extract multiple figures/tables at once
python .claude/skills/extract-pdf-figure/scripts/extract_figures.py "PDFs/paper.pdf" --batch "Figure 1,Figure 2,Figure 3,Table 1,Table 2" -d "output/asset/"

# Batch without extras
python .claude/skills/extract-pdf-figure/scripts/extract_figures.py "PDFs/paper.pdf" --batch "Figure 1,Figure 2" --no-extras -d "output/asset/"

Command Line Options

Output

• Format: PNG images
• Default location: extracted_figures/ directory next to the PDF
• Naming: <pdf_name>_<figure_name>.png

Dependencies

依赖已在项目 requirements.txt 中定义，安装方法请参考 README.md。

主要依赖：

• PyMuPDF - PDF 解析
• Pillow - 图像处理
• openai - API 客户端
• python-dotenv - 环境变量加载

How It Works

1. PDF to Image: Each PDF page is rendered at high resolution (300 DPI default)
2. AI Detection: Qwen3-VL model locates the target figure and returns bounding box coordinates
3. Quality Check: The extraction is visually assessed; if needed, coordinates are refined
4. Precise Cropping: The image is cropped with minimal padding to capture the exact figure
5. Output: Saved as an optimized PNG file

Tips for Best Results

1. Figure Names: Use exact names as they appear in the paper (e.g., "Figure 1", not "fig 1")
2. Composite Figures:
   • Use "Figure 1" to get all sub-figures with the main caption
   • Use "Figure 1(a)" to get only that specific sub-figure
3. Tables: Caption is typically above, so the tool searches upward
4. Quality: Use --dpi 400 for very detailed figures
5. Content Only: Use --no-extras when you only need the visual data without surrounding text

Additional Resources

• Script location: scripts/extract_figures.py