University of Surrey

Test tubes in the lab Research in the ATI Dance Research

Face recognition by Fisher and scatter linear discriminant analysis

Bober, M, Kucharski, K and Skarbek, W (2003) Face recognition by Fisher and scatter linear discriminant analysis Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics), 2756. pp. 638-645.

Full text not available from this repository.


Fisher linear discriminant analysis (FLDA) based on variance ratio is compared with scatter linear discriminant (SLDA) analysis based on determinant ratio. It is shown that each optimal FLDA data model is optimal SLDA data model but not opposite. The novel algorithm 2SS4LDA (two singular subspaces for LDA) is presented using two singular value decompositions applied directly to normalized multiclass input data matrix and normalized class means data matrix. It is controlled by two singular subspace dimension parameters q and r, respectively. It appears in face recognition experiments on the union of MPEG-7, Altkom, and Feret facial databases that 2SS4LDA reaches about 94% person identification rate and about 0.21 average normalized mean retrieval rank. The best face recognition performance measures are achieved for those combinations of q, r values for which the variance ratio is close to its maximum, too. None such correlation is observed for SLDA separation measure. © Springer-Verlag Berlin Heidelberg 2003.

Item Type: Article
Divisions : Surrey research (other units)
Authors :
Kucharski, K
Skarbek, W
Date : 1 December 2003
Depositing User : Symplectic Elements
Date Deposited : 17 May 2017 12:12
Last Modified : 24 Jan 2020 21:53

Actions (login required)

View Item View Item


Downloads per month over past year

Information about this web site

© The University of Surrey, Guildford, Surrey, GU2 7XH, United Kingdom.
+44 (0)1483 300800