The local discriminant bases method is a powerful algorithmic framework for feature extraction and classification applications that is based on supervised training. It is considerably faster compared to more theoretically ideal feature extraction methods such as principal component analysis or projection pursuit. In this paper, an optimization block is added to the original local discriminant bases algorithm to promote the difference between disjoint signal classes. This is done by optimally weighting the local discriminant basis using steepest decent algorithms. The proposed method is particularly useful when background features in the signal space show strong correlation with regions of interest in the signal, i.e. mammograms.