Iconclass Classification

Iconclass Classification Pipeline

A reproducible pipeline to classify digital heritage objects using Iconclass VLM via Ollama.

This project implements an automated, locally-hosted pipeline for classifying artwork images with Iconclass codes using a Vision-Language Model (VLM) running under Ollama. Designed for the Stadt Geschichte Basel digital collections, the pipeline processes metadata, downloads images, classifies them, and writes results back with full provenance tracking.

Features

🎨 Automated Iconclass Classification: Uses the Iconclass VLM model via Ollama for accurate classification
📦 Batch Processing: Process entire collections from metadata.json files
🔄 Smart Image Processing: Automatic download, resize, normalization, and SHA256-based caching
🎯 Intelligent Filtering: Automatically filters to process only children (m) objects, excluding parents (abb)
🎲 Flexible Sampling: Random, fixed, or full dataset sampling modes with reproducible seeds
📝 Multiple Prompts: Three prompt templates (default, instruction, few-shot) for optimal results
📊 Dual Output: Compact codes in metadata + detailed JSONL records for auditability
🔍 Full Provenance: Timestamped run directories with complete audit trail (requests, responses, manifests)
🛡️ Robust: Built-in retry logic, comprehensive error handling, and logging
📐 Type-Safe: Pydantic models ensure data validation and consistency
🧪 Tested: Unit and integration tests ensure reliability

Quick Start

Prerequisites

Python ≥ 3.11
Ollama running locally
The Iconclass VLM model:

ollama pull hf.co/mradermacher/iconclass-vlm-GGUF:Q4_K_M

Installation (with uv)

We use the fast Python package manager uv for reproducible environments.

# Clone the repository
git clone https://github.com/Stadt-Geschichte-Basel/iconclass-classification.git
cd iconclass-classification

# (Optional) Install uv if you don't have it yet
curl -LsSf https://astral.sh/uv/install.sh | sh

# Sync all project dependencies (creates .venv automatically)
uv sync

# Ensure dev tools (ruff, ty, pytest)
uv sync --group dev

# Show CLI help
uv run iconclass-classification --help

Basic Usage

Using Ollama (local):

uv run iconclass-classification classify-ollama \
  --source https://forschung.stadtgeschichtebasel.ch/assets/data/metadata.json \
  --model hf.co/mradermacher/iconclass-vlm-GGUF:Q4_K_M \
  --sampling-mode random \
  --sampling-size 10

Using OpenRouter (cloud):

export OPENROUTER_API_KEY=your_key_here
uv run iconclass-classification classify-openrouter \
  --source https://forschung.stadtgeschichtebasel.ch/assets/data/metadata.json \
  --model qwen/qwen3-vl-235b-a22b-instruct \
  --sampling-mode random \
  --sampling-size 10

Using a .env file for OpenRouter

You can store your API key in a local .env file and load it for the session (recommended on macOS/Linux with zsh):

# Create .env at the project root
echo 'OPENROUTER_API_KEY=your_key_here' > .env

# Load all variables from .env into the environment for this shell
set -a; source .env; set +a

# Run with OpenRouter
uv run iconclass-classification classify-openrouter \
  --source https://forschung.stadtgeschichtebasel.ch/assets/data/metadata.json \
  --model qwen/qwen3-vl-235b-a22b-instruct \
  --sampling-mode random \
  --sampling-size 10

Notes:

Keep .env out of version control. Do not commit secrets. An example.env file is provided as a template.
Alternatively, export the variable once per session using export OPENROUTER_API_KEY=....

Important: The pipeline automatically filters to only process children objects (m prefix). Parent objects (abb prefix) are excluded from classification.

The pipeline will:

Fetch metadata from the source URL
Filter out abb (parent) objects, keep only m (children) objects
Sample objects based on your configuration
Download and process images
Classify each image with the selected backend (Ollama or OpenRouter)
Write results to runs/<timestamp>/

For detailed usage instructions and advanced features, see:

USAGE.md - Complete usage guide for both backends
training-and-prompting.md - Prompt templates and troubleshooting

Output Structure

Each run creates a timestamped directory with complete provenance:

runs/<UTC-ISO8601>/
  raw/metadata.json              # Original metadata
  data/
    src/                         # Cached images (deduplicated by SHA256)
    <objectid>.jpg               # Processed images
  classify/
    <objectid>_request.json      # Classification requests
    <objectid>_response.json     # Model responses
  logs/pipeline.log              # Detailed logs
  results/
    metadata.classified.json     # Metadata with Iconclass codes
    iconclass_details.jsonl      # Detailed classification records
  manifest.json                  # Run metadata & counts

Example Output

Compact metadata (metadata.classified.json):

{
  "objectid": "abb10039",
  "title": "Die Löblich und wyt berümpt Stat Basel",
  "subject": ["71H7131", "25F2"]
}

Detailed record (iconclass_details.jsonl):

{
  "objectid": "abb10039",
  "subject": {
    "iconclass": {
      "codes": ["71H7131", "25F2"],
      "model": "hf.co/mradermacher/iconclass-vlm-GGUF:Q4_K_M",
      "image_sha256": "...",
      "timestamp": "2025-11-01T10:00:00Z"
    }
  }
}

Architecture

The pipeline is modular and consists of these components:

src/iconclass_classification/
  models.py         # Pydantic data models
  image_utils.py    # Image download & processing
  ollama_client.py  # Ollama API integration
  pipeline.py       # Pipeline orchestration
  cli.py            # Command-line interface

How It Works

Fetch Metadata: Download metadata.json from source URL
Image Selection: Choose best available image URL (prefer object_location, fallback to object_thumb)
Download & Cache: Download images with SHA256-based deduplication
Process: Resize to max size, convert to RGB, compress as JPEG
Classify: Send processed images to Ollama Iconclass VLM
Extract: Parse Iconclass codes from model responses using regex
Write: Update metadata with codes, save detailed JSONL records
Log: Save all requests, responses, and run manifest

Development

Running Tests (uv)

# Ensure dev dependencies are installed
uv sync --group dev

# Run all tests
uv run pytest test/ -v

# Run only unit tests
uv run pytest test/unit/ -v

# Run only integration tests
uv run pytest test/integration/ -v

Code Quality (uv)

# Format Python code
uv run ruff format .

# Check Python code
uv run ruff check .

# Auto-fix issues
uv run ruff check --fix .

# (Optional) Type check
uv run ty check

# Format all non-Python files (Prettier)
npm run format

# Check all non-Python files
npm run check

Troubleshooting

Ollama Not Running

Ensure Ollama is running:

ollama serve

Model Not Found

Pull the Iconclass VLM model:

ollama pull hf.co/mradermacher/iconclass-vlm-GGUF:Q4_K_M

Memory Issues

Reduce image size or use sample mode:

uv run iconclass-classification classify-ollama \
  --source <URL> \
  --model hf.co/mradermacher/iconclass-vlm-GGUF:Q4_K_M \
  --max-side 512 \
  --sample 10

Contributing

Contributions are welcome! Please see CONTRIBUTING.md for guidelines.

Support

License

Code: AGPL-3.0
Data: CC BY 4.0

Citation

If you use this project in your research, please cite:

@software{iconclass_classification,
  title = {Iconclass Classification Pipeline},
  author = {Stadt Geschichte Basel},
  year = {2025},
  url = {https://github.com/Stadt-Geschichte-Basel/iconclass-classification}
}

Acknowledgments

This project builds upon:

Iconclass classification system
Ollama for local LLM hosting
The open-research-data-template for infrastructure
Pydantic for data validation
Pillow for image processing

Authors

Stadt Geschichte Basel - Website

See also the list of contributors who participated in this project.