Online LDA which can perform both successive learning and incremental learning

摘要

Adaptability is a desired property for the computation in coming years. In particular, adaptability is important for face identification because change of situations can occur in various cases. For face identification, linear discriminant analysis (LDA) is applied extensively. However, LDA is poor at adaptability. Recently, the authors have proposed an online version of LDA, which is referred to online LDA (OLDA). By OLDA, the face identification system can be updated with low computational cost when new additional images are presented. Hence OLDA has the ability of adaptation to the change of environment. OLDA also has an advantage that N x N matrices never appear in its calculation when the number of pixels in each image is N. In the present paper, we will show experimental results that OLDA works efficiently for both two typical scenarios of the change of situations, successive learning and incremental learning.