Face Recognition Using Modular Bilinear Discriminant Analysis

We present a Modular Bilinear Disciminant Analysis (MBDA) approach for face recognition. A set of classifiers are trained independently on specific face regions, and different combination schemes are studied. The classifiers rely on a new supervised dimensionality reduction method named Bilinear Disciminant Analysis (BDA), based on a generalized bilinear projection-based Fisher criterion computed from image matrices directly. A series of experiments is performed on various international face image databases in order to evaluate and compare the effectiveness of MBDA to BDA and to the Modular Eigenspaces method. The experimental results indicate that MBDA is more efficient than both BDA and Modular Eigenspaces for face recognition.