3 Face recognition
3.3 Faces classification

Face recognition system works usually in two main phases. The first phase is a training process and the second is classification of users. Modern face recognition methods work properly when up to 10 images of one person are available in the training stage. Even the numerous techniques have been developed for face recognition from only a single image per person. Training process should be fully automated and users have to be able to control it. The training process uses clustering algorithms.

The main purpose of all clustering algorithms is to identify clusters or classes in input dataset. There are many clustering algorithms. These algorithms can be divided in two groups: partitioning and hierarchical algorithms [5].

As an example of clustering algorithm K-means can be mentioned. Another algorithm used for clustering is the self-organizing map (SOM) belonging to neural network techniques or density-based spatial clustering of applications with noise (DBSCAN).

For the classification of features extracted from faces two methods depending on the number of training images and number of identities which is to be used within the system are listed:

  • Support Vector Machines - is used when only relatively small number of identities is considered in the system. Main disadvantage of this method is the time-consuming training of the model when large number of samples is used.
  • K-Nearest neighbour distance matching (with the use of Chi-square distance) - this algorithm can be easily parallelized and used in distributed system. The training is done simply by inserting features into the database [5].