13.07.2015 Views

Thesis - Instituto de Telecomunicações

Thesis - Instituto de Telecomunicações

Thesis - Instituto de Telecomunicações

SHOW MORE
SHOW LESS

You also want an ePaper? Increase the reach of your titles

YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.

5.4. UNCERTAINTY BASED CLASSIFICATION AND CLASSIFIER FUSION 1091.01.01.01.00.80.80.80.80.60.60.60.6pe(x)0.40.40.40.40.20.20.20.20.00 5 10 15 20x=-30.00.00 5 10 15 20x=-20.00 5 10 15 20x=-10 5 10 15 20x=0Figure 5.9: Distribution of ˆp(x|w i ) at points x =[−3, −2, −1, 0].Figure 5.10: Variance of ˆp(x|w i ).to the typical rejection option using the variance of ĝ i (x ∗ ) to establish a new rejection rule.We <strong>de</strong>fine g u i (x∗ ) as:g u i (x∗ )=k[ĝ i (x ∗ )+w std(ĝ i (x ∗ ))], (5.19)where (·) <strong>de</strong>notes the mean value, and k is a normalizing factor to guarantee that ∑ n ci=1 gu i (x∗ )=1; w weights the contribution of the standard <strong>de</strong>viation of ĝ i (x ∗ ) in the modified discriminantfunction.When w → 0, g u i (x∗ ) → g i (x ∗ ), corresponding to the standard discriminant function.When w →∞, g u i (x∗ ) → std(ĝ i (x ∗ )). For our tests we selected w = 1; The new discriminantfunctions add a class <strong>de</strong>pen<strong>de</strong>nt value that will balance the uncertainty estimated for theclassification among all the classes.The proposed rejection rule is expressed by

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!