Ayelet Akselrod-Ballin 

The task of face identification is to identify an individual under variation of facial expression, viewpoint, 3D pose, illumination and occluding structure. Typically, recognition schemes involve two stages: the selection of features suitable for representation, and then the use of a classifier applied to the selected features. In this work we study the feature selection stage, in order to capture the essential information for face recognition. Selection of good features is crucial, and can lead to correct recognition, despite the variability described and in situations of highly reduced and distorted representation. The study of useful facial features was based on natural images and on artist drawings (produced by H. Piven). These drawings are reduced and often distorted representation of the real face, but are still recognizable and therefore provide evidence for features that are sufficient for efficient recognition.

We suggest that face recognition is based in part on the use of distinctive features defined as features that appear frequently in face images of a given individual but rarely in other face images. We demonstrate that when distinctive features are used it is possible to identify images from reduced and distorted representation. We show how distinctive features can be extracted for a given individual using a fragment based scheme, and test the extraction fragments on novel set of images. The results reveal that distinctive features are useful for face identification, and that only a few distinctive features are required for the reliable identification of an individual.

We demonstrate that the choice of distinctive features leads to sparse representation in contrast with compact representation often used in other schemes.

We then expand the analysis to deal with recognition as part of a hierarchy of classification tasks. We present evidence that different face features are useful for different recognition tasks (detection, identification) and present evidence that better results are obtained when features are selected according to the specific recognition requirements. The identification results are improved by applying a two-stage process of detection followed by identification, where each stage uses its own fragment, and conclude that the 2-stage process achieves better results than a single identification stage.

 

 

DownLoads

A data basse of real and piven images is available. (Piven images require permission of artist.