Weighted PCA-EFMNet: A deep learning network for Face Verification in the Wild

Abstract

The term “Wild” refers to unconstrained face recognition considered as a challenging problem due to considerable intra-class variations resulting from lighting, occlusion, facial expressions and poses changes. These challenges greatly influence the facial recognition systems performance, especially those relying on 2D information. The paper proposes an efficient deep learning network for feature extraction based on data processing components: 1) Cascaded Weighted principal component analysis with enhanced fisher model (WPCA-EFM); 2) Binary hashing; and 3) Histograms. Weighted PCA-EFM, our proposed architecture, was applied in order to learn multistage filter banks. Then, simple block histograms and simple binary hashing were applied for indexing and pooling. Therefore, the proposed architecture, named the Weighted PCA-EFM network (Weighted PCA-EFMNet), can be efficiently and easily designed and learned for Face Verification in the Wild. Ultimately, the classification is carried employing distance measure Cosine as well as support vector machine (SVM). Our experiments were carried out on real-world dataset: Labeled Faces in the Wild (LFW). Experimental results show that the proposed methods achieve high accuracy of 95%.