3D Face Analysis using Tensor Approach

Suriani, Ab Rahman (2022) 3D Face Analysis using Tensor Approach. PhD thesis, Universiti Malaysia Sarawak.

PDF (Please get the password by email to repository@unimas.my , or call ext: 082-583914/3942/3933) Suriani.pdf Restricted to Registered users only Download (3MB) | Request a copy

Abstract

The advancement of multimodal technology has enabled the creation of large face datasets. Multidimensional characteristics such as covariates and multimodal aspects in 2D, 2.5D, and 3D data are included in these datasets. Early studies in face research used matrix-based and vector-based algorithms to represent faces. According to studies, these methods have the potential to prevent the loss of critical and significant data, which could lead to lower recognition performance. The goal of this research is to develop and validate a tensor-based face recognition method that can overcome the drawbacks of matrix-based Principal Component Analysis (PCA). A face dataset consists of faces with a combination of multiple underlying causal factors such as facial expression, expression intensity, angle of view, gender and race, and the bilinear technique alone is incapable of accurately representing the dataset’s multidimensionality. PCA’s shortcomings can be overcome by using the tensor decomposition approach to separate these distinct variations. Experimental results have shown that the multilinear tensor approach could statistically outperform the bilinear PCA approach in face recognition applications. This study has added to the understanding of the centring strategy used in tensor models. The median projection operator could maximise the variation for each principal axis in the tensor space and the results have shown that in the median-centred strategy, recognition rates for emotional expressions increased from 0.4 percent to 1.4 percent. Only fear expressions have demonstrated similar recognition performance in mean-centred and median-centred experiments. Current face recognition using PCA is unable to distinguish between different types of centring approaches and cluster them. As a result, this research adopted a hybrid combined framework of a tensor model with an ANOVA (Analysis of Variance) model to uncover the substantial effects of within-subject differences (expression types and expression strengths) on recognition performance. Apart from that, experimental results revealed an interaction effect between those two covariates, showing that the effect of expression types on recognition performance is not continuous and is influenced by intensity levels. This new evidence may well be useful in a variety of recognition processes.

Actions (For repository members only: login required)

View Item