E 6886 Multimedia Security Systems Project Proposal ViewBased

E 6886 Multimedia Security Systems Project Proposal View-Based & Modular Eigenspaces for Face recognition

Team Manmohan Voniyadka Ashish Sharma Prof. Ching-Yung Lin TA Yong Wang

Goal ¡ Our proposed implementation will be based on ‘View-based & Modular Eigenspaces for Face Recognition’, Pentland et. al. proceedings of IEEE conference on computer vision and pattern recognition, 1994.

View based and Modular Eigenspaces ¡ The view-based and modular eigenspaces method for face detection and recognition is an extension of the basic eigenface method first proposed by Turk and Pentland.

Eigenface based recognition Credit: Thomas Hesltine, University of York

View based and Modular Eigenspaces ¡ The view-based formulation allows for recognition under varying head orientations and the modular description allows for incorporation of facial features. First, we calculate a view-based formulation by using separate eigenspaces for different views, each capturing the variation of all individuals in that view.

View based and Modular Eigenspaces ¡ ¡ ¡ The eigenfaces are then extended to the description of facial features to yield eigeneyes, eigenmouths or eigennoses, etc. Thus, we improve the recognition performance by incorporating an additional layer of description to the viewbased formation in terms of facial features. Extension: View-based eigenspaces for general viewing conditions Given: N individuals under M different views - Build a view-based set of M separate eigenspaces instead of a universal eigenspace from NM images.

Detection of Facial Features o In the eigenfeature representation, the equivalent Distance From Feature Space(DFFS) can be effectively used for the detection of facial features o Given an input image, a feature distance map is built by computing the DFFS at each pixel. The global minimum of this distance map is then selected as the best feature match o. The DFFS feature detection method can be extended to the detection of features under different viewing geometries with view-based eigenspaces

Feature Detection

Distance From Feature Space F DFFS DIFS F

Modular Eigenspaces ¡ With the ability to detect facial features across a wide range of faces, we can automatically generate a modular representation of a face. This layered representation is not fooled by gross variations in the input image like hats, beards etc. ¡ The topic we propose to investigate is exploring optimal fusion of the available information in modular representation. One option is to form a cumulative score in terms of equal contributions by each of the components or alternatively to give a weighting scheme in terms of the most salient features. The most robust scheme would be to have a pyramidal coarse-to-fine matching procedure to limit the search to a local region of the facespace.