Locality discriminative canonical correlation analysis for kinship verification

Metric learning has been widely used in face and kinship verification and a number of such algorithms have been proposed over the past decade. However, human face kinship datasets hold small similarity within the same sample group and obvious difference between groups, most existing methods based on single metric learning model fail to well extract features which are capable of verifying kinship relation. We apply different models for parents and children and introduce Canonical Correlation Analysis (CCA) into kinship verification which focus on multi-modal identification which can maximize the correlation between different modal data and reduces the uncertainty of the data samples, so as to achieve the purpose of enhancing the ability of recognition. Based on the advantages of CCA, we use the method of local discriminative CCA which introduces the class information of samples, and takes into account the correlation between samples influence on classification. Finally, fusion the feature extracted from parents model and the feature extracted from children model by parallel is used to complete the recognition of kinship. Experiments results demonstrate that our method is superior to some state-of-the-art methods in terms of both verification rate and computational efficiency.