Color-based feature extraction with application to facial recognition using tensor-matrix and tensor-tensor analysis

Abstract

Listen

Translate article

In this article, two main goals are investigated. 1) Developing the capabilities of tensor analysis into machine learning and pattern recognition applications such as facial recognition. 2) Answering a controversial question about the usefulness of the color information for face recognition despite small color sub-space which facial pictures possess. Firstly, an algorithm based on merely color information, extracted by tensor-matrix decomposition, is proposed. The proposed method reveals the potential of this kind of information. Secondly, another algorithm is proposed which utilizes both color and structural information by which the recognition rate is improved. Finally, by presenting the third algorithm based on tensor-tensor analysis and introducing the Eigen-Tensor concept, the task is fulfilled. The third algorithm is able to extract the color information and accomplish the dimensionality reduction simultaneously. To validate their robustness, these three proposed algorithms are evaluated by three different databases with different properties. The results represent the capabilities of tensor-tensor analysis, in comparison with other approaches. It is also proven that in spite of using the ordinary classifiers, the recognition rate more than 94% is achievable.