Enhancing performance of PCA, ICA through distribution transformation

Abstract

Holistic based face recognition methods are generally more effective on normally distributed data matrix of face images. The distribution of data matrix follows the standard Gaussian distribution according to central limit theorem, which has not been seen in practical scenarios. In this context, a simple and effective method called transformation of a data matrix to Gaussian matrix (TDG) is proposed. The TDG transforms an arbitrarily distributed data matrix to a Gaussian distribution. This transformed data matrix is then processed through Principal Component Analysis (PCA) and Independent Component Analysis (ICA). Experiments on a benchmark face database IRIS are demonstrated that the proposed transformation process could notably improve the accuracy rates than state-of-art methods.