Non-Linear Dimensional Reduction Based Deep Learning Algorithm for Face Classification

Deep learning is one of a powerful new field in Artificial intelligent, in this paper, we exhibit a technique of Autoencoder via a deep learning for Dimensionality reduction, applied on face classification. First, the Autoencoder is used to extract face features. Second, SVM kernel functions as Linear-SVM, Polynomial-SVM, Sigmoid-SVM, and Rbf-SVM are applied. Further, experiments are conducted to compare the performance of Autoencoder with two methods of dimensionality reduction which are Principal Component Analysis (PCA) and Multi-Dimensional Scaling (MDS) in order to evaluate the effectiveness of Autoencoder with the four kernel functions for face classification. Moreover, the aim of this work is to get rid of the redundant features via Autoencoder, at the same time, make a comparative study between Deep Learning method and the two previous methods, in order to test the suitability of the first one with SVM kernel functions. we evaluate the performance of our contribution using two face databases. The first one is our own created face database with hard lighting conditions named BOSS. The second one is MIT-CBCL database. The results indicate the robustness of the new proposed approach with the four kernel functions for face classification.