A novel Mutual Information based PCA approach for face identification

Principal component analysis (PCA) is a statistical tool designed to reduce dimensionality, removing redundant information in the database. When the database is a fusion of inter-similar but not intra-similar groups of databases, PCA has its limitations: poor discriminatory power and large computational load - as cumulative variance accumulation is used for selecting top features considering the global variance of the entire database. In this work, combining the information contained in the database and the Principal Components, a new approach that improves the discriminatory power of the conventional PCA is presented. This approach is developed mainly in two stages. First is the weighting of the training set through a weight vector based on mutual information. The second one is the linear projection of a weighted database using principal components. The ultimate aim of this work, given a facial image and a database of facial images, is to identify the image in the database which is most similar to the given image. The main contribution of this work is to consider the significance of the database images in the identification process in terms of the information they possess about the given image.