Character segmentation of captcha image

Question

I'm trying to crack a CAPTCHA image but I can't find a way to segment the characters.

I have this image:

I applied some filters and thresholding which resulted in this image:

Now I need to segment the image to be used in a classifier such as SVM or ANN.

The problem is that some characters are connected, and I couldn't find a way to separate them.

More image examples:

Do anyone have an approach to segment the image and get the characters?

Answer 1

It seems to me that your characters have a maximum stroke width. Whenever you find a horizontal row of black pixels that is wider than this width, that indicates that two characters are joined here.

So

• for each connected blob that is wider than a single character
  • for each row of that blob
    • find all uninterrupted horizontal lines of black pixels in this row that are longer than MAX_STROKE_WIDTH
    • note the X-coordinate of the center of these lines
  • cluster the found X-coordinates
  • split the block at the center of each cluster with more than N coordinates.
  • (you can either simply split vertically, or try to fit a line through the points in the cluster)

Answer 2

Your approach seems to be too bottom-up to me.
The number of characters is constant and they seem to be mono-spaced
So just split the image by the width and make sure the features that you retrieve are rotational/scale invariant.
This can then be fed into an ANN. I don't see why you have to segment the characters.