Feature extraction using adaptive slow feature discriminant analysis

Slow feature discriminant analysis (SFDA) is an attractive biologically inspired learning method to extract discriminant features for classification. However, SFDA heavily relies on the constructed time series. For discriminant analysis, SFDA cannot make full use of discriminant power for classification, because the type of data distribution is unknown. To address those problems, we propose a new feature extraction method called adaptive slow feature discriminant analysis (ASFDA) in this paper. First, we design a new adaptive criterion to generate within-class time series. The time series have two properties: (1) a pair of time series lies on the same sub-manifold, (2) the sub-manifold of a pair of time series is smooth. Second, ASFDA seeks projections to minimize within-class temporal variation and maximize between-class temporal variation simultaneously based on maximum margin criterion. ASFDA provides an adaptive parameter to balance between-class temporal variation and within-class temporal variation to obtain an optimal discriminant subspace. Experimental results on three benchmark face databases demonstrate that our proposed ASFDA is superior to some state-of-the-art methods.