autocrop-kh 1.1

Document Extraction Inference API using DeepLabV3 with Pretrain Model

autocrop_kh

Automatic Document Segmentation and Cropping for Khmer IDs, Passport and Documents

Autocrop_kh is a Python package for automatic document segmentation and cropping, with a focus on Khmer IDs, Passport and other documents. It uses a DeepLabV3 model training on Khmer ID, Passport document datasets to accurately segment and extract documents from images.

License: Apache-2.0 License

Installation

Install from source

# clone repo 
git clone https://github.com/MetythornPenn/autocrop_kh.git

# install lib from source
pip install -e .

Install from PyPI

pip install autocrop-kh

Usage

Python Script

import cv2
from autocrop_kh import autocrop

# Download sample image from this url : "https://github.com/MetythornPenn/autocrop_kh/raw/main/sample/img-1.jpg"
# Model auto-download (default):
# The ONNX model is downloaded on first run from Hugging Face:
# "https://huggingface.co/metythorn/autocrop/resolve/main/autocrop_model_v2.onnx"

img_path = "sample/img-1.jpg"
model_path = None

extracted_document = autocrop(
    img_path=img_path,
    model_path=model_path,
    device='cuda:0',
    output_path="extracted_document.jpg"
)

print("Extracted document saved to extracted_document.jpg")

• img_path: Path of the input image file.
• model_path: Path to the pre-trained ONNX model (local path, .onnx only). If None, it auto-downloads from Hugging Face.
• device: Specify cpu or cuda (default is cpu).
• output_path: Optional. If set, saves the extracted image to this path.
• AUTOCROP_KH_MODEL_DIR: Optional env var to change the download/cache directory.
• AUTOCROP_KH_HF_REPO: Optional env var to change the Hugging Face repo (default metythorn/autocrop).

Result:

Noted : This model was trained with 25000 datasets include opensource data and my custom synthetic data.

Reference

• Inspired by DeepLabV3