Designing A.ne Transformations Based Face Recognition Algorithms

Abstract

We investigate methods to infer the best afine transformation based face recognition algorithm; which operates by projecting given images to a low-dimensional space, followed by distance computations. This category includes the following well known methods for recognition: the Principal Component Analysis (PCA), Linear Discriminant Analysis(LDA), and Independent Component Analysis (ICA). The desired afine transformation is not restricted to that which results in an orthogonal space and can involve shear and stretch. We adopt an approach that has a reverse engineering flavor. Starting from distances computed by any face recognition algorithm, such as the FRGC baseline algorithm, we learn the best afine transform that approximates it. We propose a closed form solution for this based on classical Multidimensional Scaling (MDS). Next, this afine transform is refined by considering the modification of a given distance matrix, which will enhance the separation of match and non-match scores. The afine transform that produces the best Receiver Operating Characteristic (ROC) is selected. The data from Face Recognition Grand Challenge (FRGC-v2.0) reveals that learned afine transformation results in a better performance than the FRGC baseline algorithm.