Master Thesis - Department of Computer Science
Master Thesis - Department of Computer Science
Master Thesis - Department of Computer Science
Create successful ePaper yourself
Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.
maximum values <strong>of</strong> l and k (L and K, respectively) depend on original image size and<br />
the subspace method used.<br />
The terms Goat and Lamb originated from speaker verification [34] and then prop-<br />
agated in the field <strong>of</strong> biometry [107]. Goats are the group <strong>of</strong> subjects that are hard<br />
to authenticate and generate the majority <strong>of</strong> false rejects. The group <strong>of</strong> subjects<br />
that are easy to imitate and cause false acceptance are regarded as Lambs. Appro-<br />
priate subband for a person gathers all common intra-class detail features specific<br />
to a subject as well as eliminates the common inter-class structural features present<br />
across subjects. So, it can be stated that a proper subband face has the power to<br />
extract features that are invariant within a subject and at the same time discriminant<br />
across different subjects. Goatishness and labmishness measures for each person are<br />
calculated from the genuine and impostor distributions obtained from the confusion<br />
matrix on a validation set.<br />
Formally, let n and p be the number <strong>of</strong> samples per subject used for training and<br />
validation, respectively. The columns and rows <strong>of</strong> a confusion matrix correspond to<br />
the training and validation samples for different classes arranged sequentially. So a<br />
confusion matrix (CM) can be organized as follows,<br />
⎡<br />
m(11)(11) ⎢ . . . m(11)(1j) . . . m(11)(1n) . . . m(11)(ij) . . . m(11)(Cn)<br />
⎢ . . . . . . . . . . . . . . . . . . . . . . . . . . .<br />
⎢ m(1s)(11) . . . m(1s)(1j) . . . m(1s)(1n) . . . m(1s)(ij) . . . m(1s)(Cn)<br />
⎢ . . . . . . . . . . . . . . . . . . . . . . . . . . .<br />
⎢<br />
CM = ⎢ m(1p)(11) . . . m(1p)(1j) . . . m(1p)(1n) . . . m(1p)(ij) . . . m(1p)(Cn)<br />
⎢ . . . . . . . . . . . . . . . . . . . . . . . . . . .<br />
⎢ m(rs)(11) . . . m(rs)(1j) . . . m(rs)(1n) . . . m(rs)(ij) . . . m(rs)(Cn)<br />
⎢ . . . . . . . . . . . . . . . . . . . . . . . . . . .<br />
⎣<br />
m(Cp)(11) . . . m(Cp)(1j) . . . m(Cp)(1n) . . . m(Cp)(ij) . . . m(Cp)(Cn)<br />
(3.1)<br />
An element m(rs)(ij) represents the similarity measure (score value) between s th<br />
validation sample from class r and j th training sample from class i and takes value in<br />
the range <strong>of</strong> [0, 1]. Now the genuine and impostor score sets for class i, denoted as<br />
54<br />
⎤<br />
⎥<br />
⎦