name: px-asset-extract description: > Extract individual assets from images (slides, posters, infographics, diagrams) as transparent PNGs with a JSON manifest. Zero ML models, pure classical CV (PIL+numpy). Automatically segments, classifies (text, illustration, icon, graphic, line, dot, diagram, shadow, element), and crops each element with anti-aliased alpha transparency. Supports type filtering (--types/--exclude-types) and pre-computed region extraction (--regions) for bridging with visual grounding models. Trigger on 'extract assets from image', 'decompose slide into elements', 'get all icons from poster', 'extract illustrations', 'segment and crop', 'pull out individual elements', or when the user has an image and wants individual transparent PNGs of each element.

px-asset-extract: Image Asset Extraction

What It Does

Decomposes images into individual transparent PNG assets with classification and a JSON manifest. The full pipeline runs in 2-6 seconds on CPU with zero ML models:

1. Background detection — median color from image borders
2. Foreground mask — Euclidean color distance thresholding
3. Character bridging — dilation connects letters into words
4. Connected components — union-find with 8-connectivity
5. Classification — heuristic typing into 10 categories
6. Text-line merging — groups word fragments into text lines
7. Alpha extraction — anti-aliased transparent cropping
8. Deduplication — removes overlapping and oversized segments

When to Use This

Installation

git clone https://github.com/JadeLiu-tech/px-asset-extract.git
cd px-asset-extract
pip install .

Usage

CLI

# Basic extraction
px-extract <image> -o <output_dir>

# Only extract illustrations and icons
px-extract <image> -o <output_dir> --types illustration,icon

# Extract everything except text and dots
px-extract <image> -o <output_dir> --exclude-types text,dot,line

# Extract from pre-computed bounding boxes (e.g. from px-ground)
px-extract <image> -o <output_dir> --regions regions.json

# Segment only — output JSON, no PNGs
px-extract <image> --segments-only

# Batch processing
px-extract images/*.png -o output/ --batch

# JSON output to stdout
px-extract <image> -o <output_dir> --json --quiet

Python API

from px_asset_extract import extract_assets, load_regions

# Full extraction
result = extract_assets("slide.png", output_dir="assets/")
for asset in result.assets:
    print(f"{asset.id}: {asset.label} at ({asset.bbox.x}, {asset.bbox.y}) -> {asset.file_path}")

# Type filtering
result = extract_assets("slide.png", output_dir="icons/", types=["illustration", "icon"])
result = extract_assets("slide.png", output_dir="graphics/", exclude_types=["text", "line", "dot"])

# Pre-computed regions (from grounding model output)
regions = load_regions("grounded.json")
result = extract_assets("slide.png", output_dir="targeted/", regions=regions)

# Combine regions + type filter
result = extract_assets("slide.png", output_dir="charts/", regions=regions, types=["chart"])

CLI Options

Output

Each run produces:

• asset_NNN_<type>.png — individual transparent PNGs
• manifest.json — positions, types, and metadata for all assets
• visualization.png — input image with color-coded bounding boxes

Manifest format

{
  "source_image": "slide.png",
  "source_size": {"width": 1920, "height": 1080},
  "background_color": [255, 255, 255],
  "num_assets": 44,
  "assets": [
    {
      "id": "asset_000_illustration",
      "label": "illustration",
      "file": "asset_000_illustration.png",
      "position": {"x": 100, "y": 50, "width": 400, "height": 300},
      "pixel_area": 120000
    }
  ]
}

Regions JSON format (for --regions)

[
  {"x": 100, "y": 50, "width": 400, "height": 300, "label": "chart"},
  {"x1": 600, "y1": 100, "x2": 800, "y2": 300, "label": "logo"}
]

Also supports {"regions": [...]} wrapper. Label defaults to "region" if omitted.

Asset Types

Performance

Dependencies

Only Pillow and numpy. Optional opencv-python for better alpha edges.