Image Processing Reference
In-Depth Information
(5)
8. A binary image is obtained by using a 0.245 threshold, which was calculated experimentally
from training rainbow trout images.
9. The trout's body is emphasized by using a closing operation, first erosion, and then dilation
with a [5 × 8] mask. This operation is the key to eliminate small clusters of pixels around the
trout's body cluster.
10. The trout's contour is obtained by removing interior pixels. In this case, a pixel is set to 0 if
ifall its four-connected neighbours are 1, thus leaving only the boundary pixels on as shown
in Equation (6) :
(6)
11. Using the trout's contour, we apply our statistical measuring approach detailed in Section 3 .
12. By definition, the trout's size is estimated by computing the Mahalanobis distance from this
estimated length to training data (Equation 3 ) .
13. For classification in this experiment, imagine that the complete testing-trout set (900 in total)
is passed through a grid one by one in three steps. First, the grid is sized to filter only fry-
trout. Then, every trout able to pass this grid is labelled as a fry-trout. Second, for the rest of
the testing set, the grid is now sized to filter fingerling-trout. Every trout that is able to pass
this grid is labelled as fingerling-trout. Third, for the rest of the testing set, the grid is now
sized to filter table-fish trout.
Remember that we are computing the Mahalanobis distance and this allows us to easily im-
plement the approach above using fry, fingerling, and table-fish training data, respectively.
Referring as training data the arithmetic mean and the standard deviation from each size.
Another advantage in using Mahalanobis distance is that we can make our classification
process as rigid as we decide, by defining a threshold in number of standard deviations.
Then, in this article we are reporting classification figures from one to three standard devi-
ations.
14. We are considering this experiment as a binary classification problem, as illustrated in Table
3 . By doing this, we are collecting true positive (TP), false positive (FP), false negative (FN),
and true negative (TN) frequencies [ 18 ] .  Search WWH ::

Custom Search