6.2 Multi-class Synthetic Discriminant Functions

Synthetic discriminant functions, SDFs for short, are introduced in Chapter TM:6.1 as a way to generate a single direction onto which samples from a given set project (more or less) at the same, predefined point. The construction of an SDF relies on the possibility of solving a linear set of equations obtained by enforcing the projection values of a given set of patterns: the solution can be expressed as a linear combination of the available samples. There are two related effects that we want to explore:

1. the distribution of the projection values obtained from patterns of the same class but not explicitly used in building the SDF, and the dependency of its spread on the number of sampels used in the SDF;
2. how does an SDF compares with projection onto the (scaled) mean sample, both visually and interms of the resulting distribution of projection values.

In order to investigate the first issue, we build a sequence of SDFs, using as building samples the centers of an increasing number of clusters computed from the whole set of available faces (using function clara from package cluster). We build a least squares SDF minimizing the projection error onto the whole face dataset.

Figure 6.4: By using an increasing number of meaningful samples we are able to build an increasingly better SDF as can be seen from the progressively more peaked projection value distribution. The highly dispersed distribution closer to view, resulting from the SDF built using a single template, means that projecting onto this SDF is not going to be effective: in order to detect a good percentage of faces we need to use a high tolerance threshold, so that many false positives are to be expected. The situation is significantly better when using the SDF built from 32 samples.

The second issue can be addressed in a straightforward way: we compute the average face, the least squares SDF based on a single cluster description, and the least sqaures SDF based on a 32 cluster description:

As we can observe in Figure 6.5, the difference between sdf.m and sdf.1 is minor, but sdf.32 results in a significantly different filter providing superior performance. The dispersion of the projection values around the required value (i.e. 1) can be easily quantified by computing the variance of the values:

from which we see that there is almost an order of magnitude of difference.

Figure 6.5: SDFs provide an efficient way to compact the information from several samples into a single projection direction. The first image to the left is the average face: even if all images have been previously normalized by fixing eyes location, the remaining facial features appear blurred due to their position variability. The central image is the sample minimizing the Euclidean distance from all the other ones: is is the single cluster center returned by clara. The image to the right is the synthetic template generated by a least squares SDF built from the linear combination of 32 cluster centers.