max⎪⎧ 1⎨ −⎪⎩ 2∑( αi j yiy j Ki ) α αi ,i,j⎪⎫( x x ) ⎬ ⎪⎭jwhere,0 ≤ α ≤ C(18)∑i∑iα yiα ≥ υiiii= 0i , j ∈ 1, ...,n and the kernel function isK( x x ) = ϕ ( x ) ϕ ( x )i, (19)jij2υ -SVM is initialized and optimized using iterative decomposition training [9], whichleads to reduced complexity <strong>of</strong> 2υ -SVM. If the size <strong>of</strong> data vectors is n, then thecomplexity without optimization is O(n 2 ) and with optimization is O(n) [9]. In thissection, we consider the fusion <strong>of</strong> phase and amplitude features extracted from a faceimage. The proposed fusion algorithm, shown in Fig. 2, is described as follows:Amplitude Features<strong>Face</strong> ImageSVMLearningbasedFeature<strong>Fusion</strong>Fused FeatureVectorPhase FeaturesFig. 2. Block diagram <strong>for</strong> feature level fusion12
Let amplitude features F A and phase features F P be extracted from a face image. Both thefeatures are divided into windows <strong>of</strong> size 3 x 3 to calculate the weighted average windowbased activity. A total <strong>of</strong> n activity levels are computed and provided as input to the 2υ -SVM which is then trained to determine whether the coefficient from the amplitudefeature set F A or the phase feature set F P should be used. At any position (x, y), if thefeatures <strong>of</strong> F A are classified as “good” and the SVM classification margin <strong>of</strong> F A is greaterthan the SVM classification margin <strong>of</strong> F P , then output O(x, y) <strong>of</strong> the learning algorithm is1 otherwise the output is -1. As shown in Equation 20, amplitude and phase features areselected depending on the output O ( x,y).⎧FAFF(x,y)= ⎨⎩FP( x,y)( x,y),,ififO(x,y)> 0O(x,y)< 0(20)where FF is the fused feature vector. Furthermore, to match the two fused feature vectors,FF1and FF 2, the correlation based matching technique is applied. The features are firstdivided into m frames, each <strong>of</strong> size k x l. The correlation distance, CDiF, between twocorresponding frames is computed using Equation 21. Using the frame matchingthresholdT IF , the intermediate matching score MS IF <strong>for</strong> the frames and the finalmatching scoreMS F is calculated using Equation 22. A person is said to be matchedif MS F is greater than the fused feature vector matching threshold, T F .CDiFiiFF FF=1 ⊗ 2 , i = 1 ,2,…,m(21)k l13