Enhanced fisher discriminant criterion for image recognition

Many previous studies have shown that image recognition can be significantly improved by Fisher linear discriminant analysis (FLDA) technique. However, FLDA ignores the variation of data points from the same class, which characterizes the most important modes of variability of patterns and helps to improve the generalization capability of FLDA. Thus, the performance of FLDA on testing data is not good enough. In this paper, we propose an enhanced fisher discriminant criterion (EFDC). EFDC explicitly considers the intra-class variation and incorporates the intra-class variation into the Fisher discriminant criterion to build a robust and efficient dimensionality reduction function. EFDC obtains a subspace which best detects the discriminant structure and simultaneously preserves the modes of variability of patterns, which will result in stable intraclass representation. Experimental results on four image database show a clear improvement over the results of FLDA-based methods.