Chapter 18: Sub-tensor Decomposition for Expression Variant 3D Faces Recognition

We have investigated a technique for recognising faces invariant of facial expressions. We apply multi-linear tensor algebra, which subsumes linear algebra, to analyse and recognise 3D face surfaces. This potent framework possesses a remarkable ability to deal with the shortcomings of principle component analysis in less constrained situations. A set of vector spaces can be used to represent the variation of collections of face models with multiple formation factors across various modes, without destroying the detail of each other. Using multi-linear single value decomposition (SVD) yields better recognition rates than principal component analysis. We have used a set of landmarks as the input data for our multi-linear SVD recognition experiments. Our results have shown that the choice of landmarks may contribute to the accuracy of recognition. We have used the face action coding system (FACS) framework for manual selection of landmarks on prominent facial features as well as on muscle areas.