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 torch
import cv2
import requests
import os
from autocrop_kh import autocrop

def download_file(url, local_filename):
    with requests.get(url, stream=True) as r:
        r.raise_for_status()
        with open(local_filename, 'wb') as f:
            for chunk in r.iter_content(chunk_size=8192):
                f.write(chunk)
    print(f"Downloaded: {local_filename}")

img_url = "https://github.com/MetythornPenn/autocrop_kh/raw/main/sample/img-1.jpg"
model_url = "https://github.com/MetythornPenn/autocrop_kh/raw/main/models/autocrop_model_v2.onnx"
img_path = "img-1.jpg"
model_path = "autocrop_model_v2.onnx"

download_file(img_url, img_path)
download_file(model_url, model_path)

if not os.path.exists(img_path):
    raise FileNotFoundError(f"Image file {img_path} was not found.")
if not os.path.exists(model_path):
    raise FileNotFoundError(f"Model file {model_path} was not found.")

if torch.cuda.is_available():
    device = "cuda"
elif torch.backends.mps.is_available():
    device = "mps"
else:
    device = "cpu"

extracted_document = autocrop(
    img_path=img_path,
    model_path=model_path, 
    device=device)


output_path = "extracted_document.jpg"
cv2.imwrite(output_path, extracted_document[:, :, ::-1])  # Convert back to BGR for saving
print(f"Extracted document saved to {output_path}")

• img_path: Path of the input image file.
• model_path: Path to the pre-trained model (local path and support both .onnx and .pth).
• device: Specify cpu or cuda or mps (default is gpu).
• output_path: Path where the extracted document image will be saved.

Result:

Running as API & Web

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

# go to directory
cd autocrop

# install libraries
pip3 install -r requirements.txt

# start server (http://localhost:5555/docs)
make server

# start client ((http://localhost:7860))
make client 

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

Reference

• Inspired by DeepLabV3
• Publish python package to PyPI