An Enhanced Local Descriptor (ELD) for Face Recognition

Abstract

The influence of light changes makes the task of feature extraction more difficult for the local descriptors. Most of local descriptors sacrifices their performance in harsh lightning variations. Some uses pre-processing approach & some uses gradient based methods (with local descriptors) to improve accuracy. In this work, a novel Enhanced Local Descriptor (ELD) is introduced by taking advantages of two well behaved descriptors in harsh lightning changes. These 2 are Compound Local Binary Pattern (CLBP) & Median Robust Extended LBP based on Neighborhood Intensity (MRELBP-NI). CLBP is characterized by Sign & Magnitude details, and MRELBP-NI is characterized by Median & Mean statistics. Both of them are very essential in controlling harsh light variations. By merging features of both a discriminant descriptor ELD is gained. FLDA is taken further for size contraction & SVMs is used for matching. ELD achieves stupendous outcomes on Extended Yale B (EYB) dataset. ELD wholly outstrip the singly implemented descriptors & many methods from literature. ELD secure best accuracy of 93.42%. There is no pre-processing & the gradient based methods are used.