An Improved Face Recognition Method Using Optimized Subspace Decomposition and RBF Neural Networks

Abstract

Face biometrics is one of the most reliable tools used in applications such as personal identification and authentication, surveillance systems, and human–computer interaction. However, we usually observe the significant influence on recognition performance of these applications due to poor quality inputs affected by illumination, variations in facial pose, and overfitting challenges because of large dimensionality of image data. In order to mitigate the performance challenge, this study presents a optimized intelligent subspace decomposition-based framework to enhance the face recognition accuracy. In this framework integration of PCA, LDA, SVD, and FrSVD is used for effective feature representation and dimensionality reduction. In addition to this a weighted feature fusion mechanism is employed to integrate information obtained from the different feature extraction techniques to enhance the quality of discriminative characteristics of the resulting feature representation. The fused feature vector is subsequently classified using a Radial Basis Function (RBF) neural network. For experimentation, Yale B face data base is used to evaluate the performance of this improved framework thereby attaining the improved 96.5% of recognition accuracy compared to conventional PCA–LDA-based methods, specifically, under variations in illumination and facial pose.