Deterministic learning machine for face recognition with multi-model feature extraction

Abstract

In this paper, a new fast learning algorithm named deterministic learning machine (DLM) for the training of single-hidden layer feed-forward neural network (SLFN) subject to face recognition problem is proposed to solve the problem of high dimensional pattern recognition. The existing training algorithms for SLFN are either gradient based iterative learning algorithms or non-iterative algorithms such as extreme learning machine (ELM). The iterative learning algorithms suffer from slow learning, under-fitting, over-fitting whereas in ELM input weights are randomly chosen consequently the classification using ELM is non-deterministic. The proposed DLM is a non-iterative algorithm in which input weights are derived from input space without finding any parameter experimentally and output weights are calculated as an exact solution of linear system. This makes very fast learning and deterministic classification. The feature extraction is performed in a multi-model way by integrating the face image pixels intensity and local entropy of the image. The resulting face recognition system is highly robust against ample facial variations including illumination, pose, expression and occlusion. The proposed DLM with multi-model feature extraction is evaluated on AT&T and Yale face databases. The experimental results clearly reveal the superiority of the proposed approach.