cv-analysis-service/vidocp/redaction_detection.py

from functools import partial

import cv2
import numpy as np
import pdf2image
from iteration_utilities import starfilter, first
from matplotlib import pyplot as plt


def is_filled(hierarchy):
    # See https://stackoverflow.com/questions/60095520/how-to-distinguish-filled-circle-contour-and-unfilled-circle-contour-in-opencv
    return hierarchy[3] <= 0 and hierarchy[2] == -1


def is_boxy(contour):
    epsilon = 0.01 * cv2.arcLength(contour, True)
    approx = cv2.approxPolyDP(contour, epsilon, True)
    return len(approx) <= 10


def is_large_enough(contour, min_area):
    return cv2.contourArea(contour, False) > min_area


def is_likely_redaction(contour, hierarchy, min_area):
    return is_filled(hierarchy) and is_boxy(contour) and is_large_enough(contour, min_area)


def find_redactions(image: np.array, min_normalized_area=200000):

    min_normalized_area /= 200  # Assumes 200 DPI PDF -> image conversion resolution

    gray = ~cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)
    blurred = cv2.GaussianBlur(gray, (5, 5), 1)
    thresh = cv2.threshold(blurred, 252, 255, cv2.THRESH_BINARY)[1]

    contours, hierarchies = cv2.findContours(thresh.copy(), cv2.RETR_TREE, cv2.CHAIN_APPROX_NONE)

    contours = map(
        first, starfilter(partial(is_likely_redaction, min_area=min_normalized_area), zip(contours, hierarchies[0]))
    )
    return contours


def annotate_poly(image, contours):
    for cont in contours:
        cv2.drawContours(image, cont, -1, (0, 255, 0), 4)

    return image


def annotate_boxes_in_pdf(pdf_path, page_index=1):

    page = pdf2image.convert_from_path(pdf_path, first_page=page_index + 1, last_page=page_index + 1)[0]
    page = np.array(page)

    redaction_contours = find_redactions(page)
    page = annotate_poly(page, redaction_contours)

    fig, ax = plt.subplots(1, 1)
    fig.set_size_inches(20, 20)
    ax.imshow(page)
    plt.show()