CS 7321 Winter 1998

PS#1 Solutions by Bob Nayar

Character Recognition Problem

Index

How I solved(?) it

  1. First I converted the greyscale image into a binary image and define a colormap.
    1. This required writing my own thresholding function, because the im2bw functin stores white pixels as "1" and black pixels as "0". In order to do rotation, thinning, thickenning and erosion using the standard MATLAB functions, I had to make it a true binary image.
    2. I had to define a colormap that had white, black and the vowel colors (in that order). I also had to write a trivial function BINDIS, which incremented the binary matrix by for experimental display purposes, since the colormap is 1-based.
  2. Then I sampled each letter and "prepared" it
    1. This required "thinning" the vowel by using the 'skel' option of the BWMORPH infinite times.
  3. I then used the erode command on the binary image with the "thinned" vowel
  4. Then I found I had to do this
    1. In order to get more hits, I had to thicken the main binary image at least once.
to TOP

Assumptions and Weaknesses

I made the following assumptions
  1. The origin of the erosion element used by MATLAB's ERODE command is in the center of the erosion element. Because of this, everytime, I found a pixel in the eroded matrix, I scanned for the dimensions of the sampled vowel around that point. I don't know if this was a correct assumption or not.
  2. By "thinning" the vowel, and thickening the image, I did not have to take different samples of the vowels to get an optimal average skeleton. This was a very stupid assumption
I think the major weakness of my solutions is:
  1. By "thinning" the vowel and thickening, I introduced ambiguity up the wazoo. For example, an "o" can easily fit into the circular part of a "d" and an "i" can easily fit into an "l". This proved to be so disastrous, that I believe my program identified thickened "a"s as "e"s, because the skeleton fits into both.
  2. Being a EE, my code optimization skills are pathetic. Exercise 3 took about half an hour to run on the SUN.

to TOP

Improvements and Possible Future Work

I think that this can be improved by doing the following

to TOP

Results

Image 1

Figure 1: As seen in this picture and the other 2, I made the mistake of not changing the colormap to switch black and white.
              This picture somewhat proves my statement about letters such as "l" and "d" being falsely recognized as "i"s and "o"s.

Image 2

Figure 2: A few other observations can be made from this picture. For instance, it seems that the "o" is the easiest shape to recognize. Unfortunately, it is also obvious that by thickening the "e" and certain other letters, they are falsely identified as "o"s. Another observation is that the "game is won" by the letters that searched for last, which is why a lot of the letters are pink; ie they have been falsely identified as "u"s.

Image 3

Figure 3:  To re-iterate the last point, this is defitnitely "won" by the letter that goes last. Hence the skeleton of the "u", which was searched for last, was able to fit into almost all the thickened letters in this image, especially because all angles between 0 and 360 were tried in 5 degree increments.

to TOP

Source Code