Ambient Diagnostics - Ambient Intelligence for Scientific Discovery - page 232

reject those candidates who matched a training set of grooves. This system is
interesting for its use of a K-nearest neighbor classifier, however it also relied on
the grooves in the material to normally lie in a uniform direction. Also, the initial
portion of the detector depends on hand-tuned parameters for the morphological
operation and thresholding operations.
Crack detection is a multiple stage process that attempts to find places where
there are cracks in the tongue. Each stage is essentially a filtering process to
attempt to get rid of more unwanted information. The essential steps behind of
the process are:

1. Find all pixels that are bright in the S channel and dark in the V channel.
2. Remove pixels that fall on the edge of the tongue, as the edge area often has
the most extreme illumination variances and causes diculties for accurately
finding cracks.
3. Remove all particles in the image except the largest ones.

To convert color coordinates from RGB to HSV (Hue, Saturation, and Value),
we used the following pseudo code:

max = largest RGB component
min = smallest RGB component

60*(G-B)/(max-min) if red is largest

H= 180*(B-R)/(max-min) if green is largest

300*(R-G)/(max-min) if blue is largest

S = (max-min)/max
V = max

Step 1 is based on the discovery that cracks appear bright in the S channel,
and dark in the V channel. HSV color space is another way of representing color
image. Computers typically use the RGB color space, which separates color into
its red, green and blue components. In the RGB color space, white is represented
by red, green and blue being at maximum intensity. Purple is represented by
putting red and blue at half-intensity, and green at zero intensity. HSV represents
colors in a method more understandable by humans: H represents the hue, or
actual color, S represents the saturation, which is essentially how strong the hue
appears, and V is the overall illumination present.
The HSV color space is used to accentuate cracks. As stated before, cracks
appear bright in the S channel and dark in the V channel. “Appearing” bright
and dark, is relative to other pixels in the same channel, as they are actually
darker in the S channel, and only appear brighter because they are relatively
bright compared to their surrounding pixels. Therefore, the algorithm enhances
this contrast by thresholding local blocks relative to themselves. This is different
than a normal threshold that simply compares each pixel to a number such as
the mean of the entire image. The problem with the normal approach is that
near the edges of the tongue, there is a very high variation, which can throw