Deep Learning HMM Versus Multi-resolution transforms for Soft biometric estimation

Abstract

This work is to introduce two different methodologies to estimate the soft biometric traits from face image. The first proposed methodology to extract effective features from facial images using two multi-resolution transforms; waveatom and shearlet, for estimating gender, ethnicity, facial expression and age by Artificial Neural Network (ANN). And the second proposed methodology to use deep learning to extract suitable features by double convolutional and pooling layers to feed Hidden Markov Model (HMM) for classification. To achieve the comparative study, our experiments carried out on a large database collected from three different databases: US Adult Faces, Extended CohnKanade and FG-NET databases. The experimental results show that the multi-resolution waveatom transform was more effective than shearlet transform, but HMM with Deep learning were the best performance and more robust method to classify multi objects together such as in this paper to estimate 13 soft biometrics which clustering in four categories