K. A .D. N. K. Wimalawarne
Application of machine learning algorithms for face recognition is an active research area. In the recent past kernel methods such as Support Vector Machines (SVM) have been very successful in classifying faces. In this paper we adopt Informative Vector Machine (IVM), which is a kernel method based on Gaussian processes for face recognition. Experiments with the ORL face database has shown that recognition accuracies of both these algorithms to be comparable. But IVM has the ability to provide more sparse solutions than SVM and Autoamtic Relevance Determination (ARD) kernels were able to provide dimension reduction in feature space. Analysis of distribution of ARD values in faces has indicated a behavior that is different from the observations in component based face recognition which opens up new research ideas in face recognition. Overall, both sparse solutions and dimension reductions with IVM have reduced the storage space and computational cost while achieving recognition accuracy close to SVM.