Hierarchical Fusion of Multi Spectral Face Images for Improved ... - IIIT

training data is used, it could lead to overfitting and hence poor generalization affectingthe matching performance. Learning algorithms have several parameters which arecontrolled heuristically, making the system difficult and unreliable to use. Also,traditional multilayer neural networks suffer from the existence of multiple local minimasolutions. To alleviate these inherent limitations, we use Support Vector machine (SVM)based learning algorithms for feature fusion. SVM training always finds a globalminimum by choosing a convex learning bias. In classical SVM, the goal is to minimizethe probability of making an error. Furthermore, the parameters of SVM can be set topenalize the errors asymmetrically by assigning costs to different errors to minimize theexpected misclassification cost. This approach, known as dual υ -Support VectorMachines (2υ -SVM) [9], can also address the difficulties that arise when the classfrequencies in training data do not accurately reflect the true prior probabilities of theclasses. Considering all these properties and advantages, our proposed algorithm uses2υ -SVM [9]. 2υ -SVM is applied as a two class problem, classifying good feature classand bad feature class. Proposed by Chew et al. [9], 2υ -SVM can be expressed asfollows:x , be a set of n data vectors with xLet { } i y iid∈ R , y ∈ +1,−1and i = 1,...,n. xii is the i thdata vector that belongs to a binary class y i . The objective of training 2υ -SVM is to findthe hyper-plane that separates any two classes with the widest margins, i.e.,ω ( x)+ b = 0subject to yi( ω ϕ x)+ b) ≥ ρ −ψi( , ψ ≥ 0 ,i(14)10