inventory-ocr

This is a tool to extract structured data from fixed layout documents such as inventory cards, inventory books or tables. It requires layout definitions provided in inventory_ocr/config/regions.yaml. Currently, it is optimized to work with a collection of jewelry from the Schmuckmuseum Pforzheim, but with minor adaptations it should be adaptable to other use cases.

Specifically, to adapt this for your own data:

• Define a different document layout in inventory_ocr/config/regions.yaml
• Implement your own post-processing by subclassing the abstract PostProcessor class in inventory_ocr/postprocessor.py

Installation

Set up your environment using your favourite environment management tool. We recommend venv.

• Development (editable) install: Clone this repository, navigate into the project root and type

  pip install -e .

• Regular install: Activate your virtual environment and type

  pip install inventory-ocr

The package will install a platform-specific OCR backend by default:

• On macOS the package uses Apple Vision via the ocrmac package.
• On Windows and Linux the package uses pero-ocr.

A plain pip install will therefore pull in the appropriate OCR backend for your OS automatically (see pyproject.toml markers).

Usage

Invoke using inventory-ocr <input folder> with input folder specifying a path to the inventory cards to be processed. Results will be saved to output/ relative to the current working directory.

Command Line Arguments

inventory-ocr <input_dir> [options]

Required Arguments:

• input_dir: Path to the input directory containing files to process

Optional Arguments:

• --output_dir: Path to the output directory (default: ./output)
• --layout_config: Path to the layout configuration file (YAML). Defaults to 'inventory_ocr/config/regions.yaml'.
• --annotate: Launch the interactive layout annotation UI to create/save a regions file
• --ocr_engine: OCR engine to use (choices: auto, ocrmac, pero, mistral, dummy; default: auto)
• --eval: Run in evaluation mode (uses dummy detector and benchmarking postprocessor)

Annotation / Generating Regions

• The repository contains an example regions file under inventory_ocr/config/example_regions.yaml. You can copy this to inventory_ocr/config/regions.yaml to use it as the default layout definition.
• To create your own layout interactively, run the tool with the --annotate flag. This launches a small Gradio UI that lets you mark regions on a template image. When you press "Extract Layout" the regions are written to the layout configuration path (see --layout_config, default inventory_ocr/config/regions.yaml) and the annotation UI will close.
• If you prefer not to use the example, simply delete or rename inventory_ocr/config/example_regions.yaml, or replace inventory_ocr/config/regions.yaml with your own configuration file.

Example:

# use the bundled example regions (if present)
inventory-ocr ./data

# interactively generate and save regions to the default config path
inventory-ocr ./data --annotate

# specify a custom path for the generated regions
inventory-ocr ./data --annotate --layout_config ./my_regions.yaml

Generated YAML uses a top-level regions: mapping where each field name maps to a 4-element list [x1, y1, x2, y2] with normalized coordinates (0..1).

Supported OCR Engines

The tool supports multiple OCR engines with automatic platform-based selection:

• PERO OCR (pero): Default for non-macOS platforms
• Apple Vision API (ocrmac): Default for macOS platforms
• Mistral OCR (mistral): AI-powered OCR using Mistral models
• Dummy (dummy): For development and testing purposes
• Auto (auto): Automatically selects ocrmac on macOS, pero on other platforms

Installing Optional Dependencies

Different OCR engines require additional dependencies:

For PERO OCR (non-macOS default):

pip install inventory-ocr[pero]

For Mistral OCR:

pip install inventory-ocr[mistral]

You'll also need to obtain a Mistral API key and store it in a .env file in your project root:

1. Copy .env.template to .env:

   cp .env.template .env

2. Edit .env and replace INSERT_YOUR_KEY_HERE with your actual Mistral API key:

   MISTRAL_API_KEY=your_actual_api_key_here
   

Note: Never share your API key publicly or commit it to version control.

For Apple Vision API (macOS only): The ocrmac package is automatically installed on macOS systems. No additional installation required.

Note: The auto engine selection will use Apple Vision API on macOS (if available) and PERO OCR on other platforms.

Configuration

Region Definition

Regions (in relative coordinates) and field names are stored in inventory_ocr/config/regions.yaml (or the path you pass via --layout_config) and can be adapted according to the use case.

The region definition format uses normalized coordinates where each region is defined as [x1, y1, x2, y2] with values between 0 and 1:

• x1, y1: Top-left corner coordinates (relative to document width and height)
• x2, y2: Bottom-right corner coordinates (relative to document width and height)

custom_header_mappings: am: 'erworben am' # Maps "am" field to "erworben am" in output