Improved LBP based Descriptors in Harsh Illumination Variations For Face Recognition

Local Binary Pattern (LBP) was considered as one of the prominent local descriptors in the research community. After LBP evolution diverse LBP variants are launched in unconstrained conditions. The main problem with LBP and other descriptors are that in extreme illumination changes their performances are not adequate that's why improvements were suggested and implemented. Precisely proposed work suggests improvements to 3 local descriptors i.e. LBP, Horizontal Elliptical LBP (HELBP) and Median Binary Pattern (MBP). The improvements were suggested by deploying 2 Dimensional-Discrete Wavelet Transform (2D-DWT) prior to feature extraction. After employing 2D-DWT (utilizing haar at level 1), the input image is decomposed into 4 sub-bands. First one signifies approximation coefficient and rest three signifies detail coefficients which are horizontal, vertical and diagonal. Then LBP, HELBP and MBP histograms were separately extracted from 4 wavelet sub-bands. The sub-band histograms were fused for developing the respective descriptor feature length. These 3 improved descriptors are defined as 2D-DWT+LBP, 2D-DWT+HELBP and 2D-DWT+MBP. Fishers Linear Discriminant Analysis (FLDA) was used for reducing the dimension. Then classification was attained by the Radial Basis Function (RBF) (the Support Vector Machines (SVMs) based method). On Yale B (YB) and Extended YB (EYB) datasets, improved descriptors beats the outcome of original descriptors completely.