Image Processing Reference
Precision measures that fraction of examples classified as positive that are truly positive:
5 Performance evaluation
We now present performance figures when using our statistical model to measure rainbow
trout in a farm.
As observed in Figure 5 , our statistical approach's performance to measure a rainbow trout
depends on our image processing stage. However, according to our experimental results, we
believe that we have addressed main issues about capturing and processing an RGB image
within our system.
As we have mentioned before, we consider this as a binary classification problem as indic-
(computed from 900 images) are compared against fry training data, using Mahalanobis dis-
tance. Then, if a testing length falls into a predefined threshold (one to three standard devi-
ations), the respective testing trout is marked as fry-trout. Next, all remaining testing lengths
are compared against fingerling training data using Mahalanobis distance as well. Again, if
the testing length falls into a predefined threshold (one to three standard deviations), we all
bel the respective testing trout as fingerling-trout. Then, every remaining testing length is
compared against table-fish training data using Mahalanobis distance. Similarly, if the testing
length falls into a predefined threshold (one to three standard deviations), the testing trout is
labelled as table-ish-trout.
Then, as prescribed in Table 3 we count TP, FP, TN, and FN frequencies, which are sum-
marized from Tables 4 to 6 . Hence, by using these values we are able to compute accuracy,
repeatability, and specificity metrics, which are presented from Tables 7 to 9 and illustrated in