PSO versus AdaBoost for feature selection in multimodal biometrics

In this paper, we present an efficient feature level fusion scheme that we apply on face and palmprint images. The features for each modality are obtained using Log Gabor transform and concatenated to form a fused feature vector. We then use Particle Swarm Optimization (PSO) scheme to reduce the dimension of this vector. Final classification is performed on the projection space of the selected features using Kernel Direct Discriminant Analysis (KDDA). Extensive experiments are carried out on a virtual multimodal biometric database of 250 users built from the face FRGC and the palmprint PolyU databases. We compare the proposed selection method with the well known Adaptive Boosting (AdaBoost) method in terms of both number of features selected and performance. Experimental results in both closed identification and verification rates show that feature fusion improves performance over match score level fusion and also that the proposed method outperforms AdaBoost in terms of reduction of the number of features and facility of implementation.