Supervised feature extraction for tensor objects based on maximization of mutual information

• We propose a method for supervised feature extraction for tensor objects.
• Features are extracted by maximizing an approximation of mutual information.
• The objective function uses information beyond the second order statistics.
• Experiments justify additional complexity with a clear performance improvement.

Several supervised feature extraction methods for tensor objects have been proposed recently, with applications in recognition of objects, faces and handwritten digits. However, the existing methods usually use only second order statistics of the data, typically through calculation of the within- and between-class scatters. Here we propose a method for supervised feature extraction for tensor objects based on maximization of an approximation of mutual information. In this way we utilize information contained in the higher order statistics of the data. Several experiments show that the proposed method results in highly discriminative features.