Image Processing Reference
Fig. 13.7 The feature space
and the distribution of the 60
coins used for training
“I have an idea,” SB said, “if the center pixel is white we don't have a hole. If
the center pixel is black we start a connected component analysis in this pixel where
object pixels are black. If the BLOB found in this way has a size corresponding to
the size of a hole and ... ”. SB stopped when he noticed Mick wasn't paying attention
“How can a hole have a size?” Mick said with a wolfish smile.
“Idiot,” SB said and continued, “if the size of the BLOB is acceptable we have
found a hole.”
“But a hole is something that is not there, so how can you then find it?” laughed
Mick, while SB left the room to get some air. When he returned they finished im-
plementing the hole detection algorithm. Next they looked at the feature space, see
Fig. 13.7 .
“Why can a certain type of coin have different sizes from image to image? I mean,
a certain type of coin has a fixed size” SB wondered aloud.
“First of all, the further away from the camera a coin is the smaller it will look.”
“Really?” SB asked doubtfully.
“Yes. Imagine a coin 10 m away, then it would be very small in the image.”
“Ahh, I get it, that is due to the perspective geometry of the camera.” Mick looked
really impressed at SB while wondering where he picked up such a fancy term.
“Second, and more importantly,” Mick continued, “due to changes in the light-
ing and the different surfaces of different coins, the segmentation will not perform
equally well on different coins and hence the area of even the same type of coins
will change. But if you look at the feature space and focus on the coins with a hole,
you can see that there is a relatively large difference between the sizes of the differ-
ent types. From this follows that we can simply classify the coins without a hole by
a few if-then-else statements. You can see the same is true for the coins with a hole.
This classification strategy gives us the type of a particular BLOB. The center of a
coin we can find using Eq. 7.2. And no, we can't use one of the approximations of
the center like last time, since we need as precise a center as possible.”
They implemented this and the resulting information sent to the robot was as in
Table 13.1 .
The last step was to test their algorithm and they therefore captured a number of
images in different lighting conditions and with different coins located at different
locations. They found that sometimes the lighting resulted in poor segmentation