DWT and LBP Map Based Feature Descriptors for Face Recognition in Harsh Light Variations

Abstract

In [11] Karanwal et al. provided enhancements to three descriptors for Face Recognition under illumination variations. The three descriptors for which enhancements are done are LBP, HELBP and MBP. By deploying 2D-DWT (utilizing haar at level 1) before features extraction of LBP, HELBP and MBP, the enhancements are made. These improved ones outperforms the original descriptors comprehensively. After careful analyzing the work proposed in [11] it has been observed that even after image pre-processing, histograms of LBP, HELBP and MBP unable to capture the efficient information to declare as the robust descriptors in light variations. In the proposed work it has been observed and implemented that map feature of LBP, HELBP and MBP (after image pre-processing by 2D-DWT) yields much better accuracy than the histogram based descriptors. The three proposed descriptors are 2D-DWT+LBPmap, 2D-DWT+HELBPmap, 2D-DWT+MBPmap. These map features full and completely outperform its respective histogram features & these are LBPhist, 2D-DWT+LBPhist, HELBPhist, 2D-DWT+HELBPhist, MBPhist and 2D-DWT+MBPhist. Among all it is 2D-DWT+HELBPmap feature which yields best results. The feature compression is fulfilled by the usage of Fishers Linear Discriminant Analysis (FLDA) and classification was done from Support Vector Machines (SVM)s. For experiments Yale B (YB) and Extended Yale B (EYB) datasets are used.