app.py

# Import the required libraries
import gradio as gr
import cv2 # OpenCV, to read and manipulate images
import easyocr # EasyOCR, for OCR
import torch # PyTorch, for deep learning  
import pymupdf # PDF manipulation
from transformers import pipeline # Hugging Face Transformers, for NER
import os # OS, for file operations
from glob import glob # Glob, to get file paths

##########################################################################################################
# Initiate the models

# Easyocr model
print("Initiating easyocr")
reader = easyocr.Reader(['en'], gpu=torch.cuda.is_available(), model_storage_directory='.')

# Use gpu if available
print("Using gpu if available")
device = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu")
print(f"Using device: {device}")

# Ner model
print("Initiating nlp pipeline")
nlp = pipeline("token-classification", model="dslim/distilbert-NER", device=device)

##########################################################################################################
## Functions

# Define img_format
img_format = "png"

# Convert pdf to set of images
def convert_to_images(pdf_file_path):

    # Create a directory to store pdf images
    pdf_images_dir = f'{pdf_file_path}_images'
    os.makedirs(pdf_images_dir, exist_ok=True)

    # DPI
    dpi = 150

    # Convert the PDF to images
    print("Converting PDF to images...")
    doc = pymupdf.open(pdf_file_path)  # open document
    for page in doc:  # iterate through the pages
        pix = page.get_pixmap(dpi=dpi)  # render page to an image
        pix.save(f"{pdf_images_dir}/page-{page.number}.{img_format}")  # store image as a PNG

    # Return the directory with the images
    return pdf_images_dir

# Do the redaction
def redact_image(pdf_image_path, redaction_score_threshold):

    # Loop through the images
    print("Redacting sensitive information...")

    print(f"Processing {pdf_image_path}...")
    # Read the image
    cv_image = cv2.imread(pdf_image_path)

    # Read the text from the image
    result = reader.readtext(cv_image, height_ths=0, width_ths=0, x_ths=0, y_ths=0)

    # Get the text from the result
    text = ' '.join([text for (bbox, text, prob) in result])

    # Perform NER on the text
    ner_results = nlp(text)

    # Draw bounding boxes
    for ((bbox, text, prob),ner_result) in zip(result, ner_results):

        # Get the coordinates of the bounding box
        (top_left, top_right, bottom_right, bottom_left) = bbox
        top_left = tuple(map(int, top_left))
        bottom_right = tuple(map(int, bottom_right))

        # Calculate the centers of the top and bottom of the bounding box
        # center_top = (int((top_left[0] + top_right[0]) / 2), int((top_left[1] + top_right[1]) / 2))
        # center_bottom = (int((bottom_left[0] + bottom_right[0]) / 2), int((bottom_left[1] + bottom_right[1]) / 2))


        # If the NER result is not empty, and the score is high
        if len(ner_result) > 0 and ner_result['score'] > redaction_score_threshold:

            # Get the entity and score
            # entity = ner_result[0]['entity']
            # score = str(ner_result[0]['score'])

            # Apply a irreversible redaction
            cv2.rectangle(cv_image, top_left, bottom_right, (0, 0, 0), -1)
        # else:
            # entity = 'O'
            # score = '0'
            
        # # Draw the bounding box
        # cv2.rectangle(cv_image, top_left, bottom_right, (0, 255, 0), 1)
        # # Draw the entity and score
        # cv2.putText(cv_image, entity, center_top, cv2.FONT_HERSHEY_SIMPLEX, 0.5, (0, 0, 255), 2)
        # cv2.putText(cv_image, score, center_bottom, cv2.FONT_HERSHEY_SIMPLEX, 0.5, (255, 0, 0), 2)

    # Save the redacted image
    print(f"Saving redacted {pdf_image_path}...")
    redacted_image_path = pdf_image_path.replace(f'.{img_format}', f'_redacted.{img_format}')
    # Save the redacted image in png format
    cv2.imwrite(redacted_image_path, cv_image)

    return redacted_image_path

# Convert the set of redacted images to a pdf
def stich_images_to_pdf(redacted_image_files, input_pdf_path):

    # Sort the redacted images
    redacted_image_files.sort()

    # Convert the redacted images to a single PDF
    print("Converting redacted images to PDF...")
    redacted_pdf_path = input_pdf_path.replace('.pdf', '_redacted.pdf')

    doc = pymupdf.open()
    for redacted_image_file in redacted_image_files:
        img = pymupdf.open(redacted_image_file)  # open pic as document
        rect = img[0].rect  # pic dimension
        pdfbytes = img.convert_to_pdf()  # make a PDF stream
        img.close()  # no longer needed
        imgPDF = pymupdf.open("pdf", pdfbytes)  # open stream as PDF
        page = doc.new_page(width = rect.width,  # new page with ...
                        height = rect.height)  # pic dimension
        page.show_pdf_page(rect, imgPDF, 0)  # image fills the page
    doc.save(redacted_pdf_path)

    # print(f"PDF saved as {redacted_pdf_path}")

    return redacted_pdf_path

def cleanup(redacted_image_files, pdf_images, pdf_images_dir, original_pdf):

    # Remove the directory with the images
    print("Cleaning up...")

    # Remove the redacted images
    for file in redacted_image_files:
        os.remove(file)

    # Remove the pdf images
    for file in pdf_images:
        os.remove(file)

    # Remove the pdf images directory
    os.rmdir(pdf_images_dir)

    # Remove original pdf
    os.remove(original_pdf)

    return None

# Func to control ui
def predict(input_pdf_path, sensitivity):

    print("Setting threshold")
    # Convert sensitivity to threshold
    redaction_score_threshold = (100-sensitivity)/100

    # Convert the PDF to images
    print("Converting pdf to images")
    pdf_images_dir = convert_to_images(input_pdf_path)

    # Get the file paths of the images
    print("Gathering converted images")
    pdf_images = glob(f'{pdf_images_dir}/*.{img_format}', recursive=True)
    pdf_images.sort()

    # Redact images
    print("Redacting images")
    redacted_image_files = []

    for pdf_image in pdf_images:

        redacted_image_files.append(redact_image(pdf_image, redaction_score_threshold))


    # Convert the redacted images to a single PDF
    print("Stitching images to pdf")
    redacted_pdf_path = stich_images_to_pdf(redacted_image_files, input_pdf_path)

    print("Cleaning up")
    cleanup(redacted_image_files, pdf_images, pdf_images_dir, input_pdf_path)

    return redacted_pdf_path

##########################################################################################################

contact_text = """
# Contact Information

👤  [Mitanshu Sukhwani](https://www.linkedin.com/in/mitanshusukhwani/)

✉️  mitanshu.sukhwani@gmail.com

🐙  [mitanshu7](https://github.com/mitanshu7)
"""

##########################################################################################################
# Gradio interface

with gr.Blocks(theme=gr.themes.Soft()) as demo:

    # Title and description
    gr.Markdown("# RedactNLP: Redact your PDF!")
    gr.Markdown("## How redaction happens:")
    gr.Markdown("""
                1. The PDF pages are converted to images using **[PyMuPDF](https://github.com/pymupdf/PyMuPDF)**.
                2. **[EasyOCR](https://github.com/JaidedAI/EasyOCR)** is run on the converted images to extract text.
                3. **[dslim/distilbert-NER](https://huggingface.co/dslim/distilbert-NER)** model does the token classification.
                4. Non-recoverable mask is applied to identified elements using **[OpenCV](https://github.com/opencv/opencv)**.
                5. The masked images are converted back to a PDF again using **[PyMuPDF](https://github.com/pymupdf/PyMuPDF)**.
                """)
    gr.Markdown("*Note: If you already have a ML setup, it is preferable that you download the [github repo](https://github.com/mitanshu7/RedactNLP) and use it offline. It offers better privacy and can use GPU for (much) faster computations while utilising a better model like **[FacebookAI/xlm-roberta-large-finetuned-conll03-english](https://huggingface.co/FacebookAI/xlm-roberta-large-finetuned-conll03-english)** or **[blaze999/Medical-NER](https://huggingface.co/blaze999/Medical-NER)***")
    
    # Input Section
    pdf_file_input = gr.File(file_count='single', file_types=['pdf'], label='Upload PDF', show_label=True, interactive=True)
    
    # Slider for results count
    slider_input = gr.Slider(
        minimum=0, maximum=100, value=80, step=1, 
        label="Sensitivity to remove elements. Higher is more sensitive, hence will redact aggresively."
    )

    # Submission Button
    submit_btn = gr.Button("Redact")
    
    # Output section
    output = gr.File(file_count='single', file_types=['pdf'], label='Download redacted PDF', show_label=True, interactive=False)

    # Attribution
    gr.Markdown(contact_text)

    # Link button click to the prediction function
    submit_btn.click(predict, [pdf_file_input, slider_input], output)


################################################################################

if __name__ == "__main__":
    demo.launch()