Removing circles, lines, and other patterns with OpenCV and Python, part 2

It took me a while to come up with an (at least) semi-effective method of removing the text pattern. While the thicker outlines can be disposed of pretty easily (click, click), developing a method which wouldn’t harm the text too much turned out to be much more challenging.

import matplotlib.pyplot as plt
import numpy as np
import cv2
%matplotlib inline

# a little helper function to display our image in a bigger plot
def display_img(image):
    fig = plt.figure(figsize=(20,16))
    ax = fig.add_subplot(111)
    ax.imshow(image, cmap="gray")
img = cv2.imread('obfuscated.jpg', 0)
img_med = cv2.medianBlur(img, ksize=15)
ret, th = cv2.threshold(img_med, 80, 255, cv2.THRESH_BINARY)
kernel = np.ones((3,3),np.uint8)
eroded_img = cv2.erode(th,kernel,iterations=2)
img_dest = img.copy()
img_dest[eroded_img == 0] = 255
display_img(img_dest)

png

First, let’s remove the obfuscation from the margins by simply changing the colour of all the margin pixels to white.

img_dest[:, :900] = 255
img_dest[:, 3350:] = 255
img_dest[:900, :] = 255
img_dest[4200:, :] = 255
display_img(img_dest)

png

Next — erode the image with a horizontal kernel to get some “FBI-redacted” style text.

kernel = np.ones((6,16),np.uint8)
eroded = cv2.erode(img_dest,kernel,iterations=6)
ret, th = cv2.threshold(eroded, 127, 255, cv2.THRESH_BINARY)
display_img(th)

png

As you can observe, the iritating vertical blobs are the remaining “obfuscated” tags. At this point I hoped I could get something out of this by measuring the gradient directions, but it would take way too much text away with it… so custom methods it is.

First, if in any row there is less black than my expected threshold, then the whole row can go to hell.

for index, row in enumerate(th):
    if cv2.countNonZero(row) > 0.6*len(row):
        th[index] = np.uint8(np.repeat(255,len(row)))
display_img(th)

png

Second, to only leave the longest sequence in the row let’s first dilate it even more. (The kernel size depends on the line height — if it’s bigger and if there’s more space between lines you can afford adding a bit more in the vertical axis. In this case it would turn the text into very thin lines.)

kernel = np.ones((2,8),np.uint8)
dilated = cv2.dilate(th,kernel,iterations=7)
display_img(dilated)

png

img_dest[dilated == 255] = 255
display_img(img_dest)

png

Again, depending on the exact document, you may have to use different parameters with the morphological operators, and the effect will be more (or less) prominent.

Published 14 Apr 2019

Personal site; contains my incoherent ramblings, with trace amounts of code & dry jokes.
MK Mozgawa on Twitter