Blur Face Recognition using Blur Metric and Some Variants of Supervised Distance Preserving Projection

Abstract

This paper is focused on face recognition techniques in uncontrolled scenarios, specifically on the recognition of face images with blur effects. At first, the blur level of the testing image is determined using recently proposed blur metric. This blur metric value is used to blur the training set of gallery images using Gaussian filter. The blur level of training images is the same as that of the testing image. Two variants of Supervised Distance Preserving Projection (SDPP), SDPP as Semidefinite Least Square (SLS-SDPP) and Regularized Supervised Distance Preserving Projection (RSDPP), are used for extracting effective features of training and testing images. K-Nearest Neighbor classifier is used for matching. Numerical experiments were carried out on two benchmarking face data ORL and Yale. The performances of SLS-SDPP and RSDPP are compared with one of the leading methods Eigenface method. Experimental results show that the combination of blur metric and the feature extraction methods achieved outstanding performance in recognizing blur images of different levels and also outperforms the base methods and Eigenface method.

Dhaka Univ. J. Sci. 69(3): 154-160, 2022 (June)