Modern CRAFT Library

Overview

This Python library provides an interface for using the CRAFT (Character Region Awareness For Text Detection) algorithm for text detection in images. The library is based on the CRAFT implementation in Python.

This is an adaptation of the CRAFT-python library allowing for use on modern packages (pytorch 2 > opencv-python > 3.4).

GPU (CUDA) Acceleration has now been tested.

ToDo

This library will need refactoring with respect to performance as the previous library introduced unneeded overhead.

Requirements

• Python 3.x
• PyTorch
• OpenCV
• NumPy

Installation

You can install the library via pip:

pip install mcraft

Usage

    from mcraft import TNet
    
    # Load an image
    image_path = "path/to/your/image.jpg"
    
    # Initialize the network
    tnet = TNet()  # Adjust parameters as needed
    
    # Run a test
    tnet.test(image_path)
    
    # Perform text detection
    tnet.test(image_path)
    
    # After running test, the result will be saved as "res_image_mask.jpg" in the same directory as the input image

Features

• Text detection in images produced as heat-maps bounding boxes and per-character polygons.
• Adjustable parameters such as text threshold, link threshold, etc.

Important Notes

• Preexisting or hand compiled instances of opencv, pytorch, and torchvision will not be replaced during setup.
• The necessary model files (craft_mlt_25k.pth and craft_refiner_CTW1500.pth) will be downloaded to ~/.cache/mcraft.
• GPU operations, whilst implemented, remain untested until I get back from abroad and regain access to my compute systems.
• Official CRAFT PyTorch repository

License

This project is licensed under the MIT License - see the LICENSE file for details.