bistr-o-mathik.org

We now have these nifty copying machines that allow you to save scanned pages to a PDF document instead of printing them. Using Acrobat Pro, it is very easy to turn these scans into fully searchable PDFs. Much better than simple paper copies. So, last time I wanted to keep some pages of a book for reference, guess what I did.

Unfortunately, after returning the book to the library I discovered that the default settings of the copier try to save space and produce heavily pixelated black and white images. The text is nearly unreadable. Zooming in to 200%, the scan looks like this:

Nevertheless, the problem can be fixed using morphological operations which are readily available in OpenCV. After applying a closing operation (a dilation followed by an erosion) with a small circle as structuring element the image already looks much better (pixels added by the closing highlighted in red):

As you can see, the closing operation fills all the tiny gaps but keeps the shape of letters intact. Finally, I applied a small blur filter and then increased the contrast to get nice, smooth edges. Here is a comparison of the scan before and after the repair:

Now, that is something I can live with. And all of this can be done in just a few lines of code:

#!/usr/bin/env python
import sys
import cv2
import numpy as np
 
kernel = cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (3, 3))
 
for filename in sys.argv[1:]:
    print "Processing {0:s}n".format(filename)
 
    # read image as grayscale:
    img = cv2.imread(filename, 0)
 
    # close gaps:
    cv2.erode(img, kernel, img)
    cv2.dilate(img, kernel, img)
 
    # blur and then increase contrast to get smooth edges:
    img = cv2.GaussianBlur(img, (5, 5), 0)
    tmp = (img.astype('int32') - 25) * ((255 + 50) / 255.0)
    np.clip(tmp, 0, 255, out=tmp)
    img = tmp.astype('uint8')
 
    cv2.imwrite(filename, img)

OpenCV’s drawing functions for geometric primitives like lines and circles regrettably do not support alpha-transparency. As workaround, the documentation remarks that one should paint to a separate buffer and then blend this buffer with the main image. But how does this work in practise?

A convenient way to blend images is via the addWeighted function (there is an explanation of image overlays on AI Shack that basically reimplements addWeighted by hand). To get standard transparency, simply make sure the weights add up to 1. Here’s the code I used to create the image above:

#!/usr/bin/env python
import cv2
 
img = cv2.imread('lena.png')
 
# (1) create a copy of the original:
overlay = img.copy()
# (2) draw shapes:
cv2.circle(overlay, (133, 132), 12, (0, 255, 0), -1)
cv2.circle(overlay, (166, 132), 12, (0, 255, 0), -1)
# (3) blend with the original:
opacity = 0.4
cv2.addWeighted(overlay, opacity, img, 1 - opacity, 0, img)
 
# display result (press 'q' to quit):
cv2.namedWindow('Transparency')
cv2.imshow('Transparency', img)
while (cv2.waitKey() & 0xff) != ord('q'): pass
cv2.destroyAllWindows()

What’s important about this method is to start the overlay buffer with a copy of the original, not with a blank image. With this neat little trick, all points that are not affected by the drawing operations will retain their color. Otherwise, they are blended with the background color of the overlay buffer.