How can I blue-box an image?

Question

I have a scanned image which is basically black print on some weird (non-gray) background, say, green or yellow (think old paper).

How can I get rid of the green/yellow and receive a gray picture with as much of the gray structure of the original image intact? I.e. I want to keep the gray around the letters for the anti-aliasing effect or for gray areas but I want to turn anything which even is remotely green/yellow to become pure white?

Note that the background is by no means homogeneous; so the algorithm should be able accept a color and an error margin or a color range.

For bonus points: How can I automatically determine the background color?

I'd like to use Python with the Imaging Library or maybe ImageMagick.

Note: I'm aware of packages like unpaper. My problem with unpaper is that it produces B&W images which probably look good for an OCR software but not for the human eye.

Answer 1

I am more of C++ than python programmer, so I can't give you a code sample. But the general algorithm is something like this:

Finding the background color: You make a histogram of the image. The histogram should have two peaks representing the background and foreground colors. Because you know that the background has higher intensity you choose the peak with higher intensity and that is the background color. Now you have the RGB background (R_bg, G_bg, B_bg)

Setting the background to white: You loop over all pixels and calculates the distance to the background:

distance = sqrt((R_bg - R_pixel) ^ 2 + (G_bg - G_pixel) ^ 2 + (B_bg - B_pixel) ^ 2)

If the distance is less than a threshold you set the pixel to white. You can experiment with different thresholds until you get a good result.

Answer 2

I was looking to make an arbitrary background color transparent a while ago and developed this script. It takes the most popular (background) color in an image and creates an alpha mask where the transparency is proportional to the distance from the background color. Taking RGB colorspace distances is an expensive process for large images so I've tried some optimization using numpy and a fast integer sqrt approximation operation. Converting to HSV first might be the right approach. If you havn't solved your problem yet, I hope this helps:

from PIL import Image
import sys, time, numpy

fldr = r'C:\python_apps'
fp = fldr+'\\IMG_0377.jpg'

rz = 0  # 2 will halve the size of the image, etc..

# ----------------

im = Image.open(fp)

if rz:
    w,h = im.size
    im = im.resize((w/rz,h/rz))
    w,h = im.size

h = im.histogram()
rgb = r0,g0,b0 = [b.index(max(b)) for b in [ h[i*256:(i+1)*256] for i in range(3) ]]

def isqrt(n):
    xn = 1
    xn1 = (xn + n/xn)/2
    while abs(xn1 - xn) > 1:
        xn = xn1
        xn1 = (xn + n/xn)/2
    while xn1*xn1 > n:
        xn1 -= 1
    return xn1

vsqrt = numpy.vectorize(isqrt)

def dist(image):
    imarr = numpy.asarray(image, dtype=numpy.int32)  # dtype=numpy.int8
    d = (imarr[:,:,0]-r0)**2 + (imarr[:,:,1]-g0)**2 + (imarr[:,:,2]-b0)**2
    d = numpy.asarray((vsqrt(d)).clip(0,255), dtype=numpy.uint8)
    return Image.fromarray(d,'L')

im.putalpha(dist(im))
im.save(fldr+'\\test.png')