Unconstrained and constrained face recognition using dense local descriptor with ensemble framework

This paper presents an ensemble face recognition system which makes use of the novel local descriptor called Dense Local Graph Structure (D-LGS) which is exploited from symmetric LGS that and it uses additional graph structure in addition to its own local graph structure. This additional local graph structure is generated by finding additional corner pixel points through bilinear interpolation of neighborhood pixels. These corner pixels lead to most stable features and information related to local deformation of the image. In this proposed ensemble system, three classifiers, namely K-nearest neighbor, Chi-square and correlation coefficient are used. Further the proposed approach fuses the decisions obtained from individual classifiers through OR rule, majority voting and AND rule. To evaluate the performance of proposed ensemble system, the experiment is conducted with three face databases viz. AT&T (formerly The ORL Database of Faces), UFI and LFW face database. The ensemble face recognition system on the use of novel dense local graph structure has reached the accuracy of 100% on AT&T, 99.3488% on UFI and 87.3372% on LFW face database. Further, the templates of D-LGS are optimized using Genetic algorithm (GA) as part of 'curse-of-dimensionality' and the reduced number of templates give accuracies of 100% on AT&T and 99.2165% on LFW face database. (c) 2020 Elsevier B.V. All rights reserved.