1

I am new in matlab and I am trying to write a code that can separate hand written numbers form a squared page background. I have tried using kmeans to differ the numbers from the lines and the squers , but it doesn’t work on all images . Any ideas how can I do it? Tahnks!

Shai
  • 111,146
  • 38
  • 238
  • 371
user2339198
  • 69
  • 2
  • 6
  • Could you show an example image + the result of your attempt (including the used code)? – Dennis Jaheruddin May 01 '13 at 11:52
  • Sure, here is a link to download the code : http://www.speedyshare.com/9KAK7/Untitled.m and a link to a "good" picture : http://www.speedyshare.com/X8c8T/1.jpeg and link to a "bad" picture : http://www.speedyshare.com/RCSaU/badpic.jpg – user2339198 May 01 '13 at 12:19
  • and here is a word document that shows the results : http://www.speedyshare.com/B3rkx/results.docx – user2339198 May 01 '13 at 12:26
  • Interesting project you have here. While there are many complex computer vision and image processing solutions to this problem, I think your solution so far is pretty good as far as simple implementations go. I think one technique to look at would be 2D fourier tranforms; the periodic grid lines in the background are a prime target for frequency filters. – Suedocode May 02 '13 at 00:10
  • ur links don't work anymore. Could you post workable links again? I'd like to have a loot at it. Thanks. – magarwal Oct 11 '13 at 13:35

1 Answers1

2

I'm using the EBImage package in R, but I'm sure you can find the equivilent in matlab:

Starting with the original image: enter image description here

# Read and extract greyscale image 
# Run kmeans with 3 centers 
km = kmeans(as.vector(x), 3)
c = km$centers

# 1 0.4936797
# 2 0.2841005
# 3 0.6456494

# Extract smallest cluster 2 (your numbers) as binary image
t = (km$cluster==2)
r = matrix(t, nrow(x))

r now looks like:

enter image description here

# Apply morphological opening (erode then dilate) with smallest possible structural element
kern = makeBrush(3, 'box')

#[,1] [,2] [,3]
#[1,]    1    1    1
#[2,]    1    1    1
#[3,]    1    1    1

z = openingGreyScale(r, kern)

z now looks like

enter image description here

Omar Wagih
  • 8,504
  • 7
  • 59
  • 75