Face Recognition with Local High-Order Principal Direction Pattern Based on “Gradient Face”

Abstract

Pointing to weak robustness caused by the noise sensitivity and feature redundancy of present face recognition methods with high-order features, a new method of the local high-order principal direction pattern based on “gradient face” is proposed. Firstly, the gradient face convolution operator designed is used to compute the sum of multi-directional gradient components of pixels to construct a gradient face. Then, the principal direction grouping strategy is introduced on the gradient face to characterize its high-order derivative features, and a principal direction feature map is formed according to the feature code of high-order derivatives direction changes in local neighborhood. Finally, block statistics and cascading of histogram features are made a vector to be input in to a support vector machine for multi-classification. Experimental results of several public face databases show that the proposed method is robust to changes in illumination, expression, and facial occlusion and has higher recognition efficiency.