Kinship verification using multi-level dictionary pair learning for multiple resolution images

Abstract

Kinship verification using facial images is gaining substantial attention by computer vision researchers. The real challenge in kinship verification is to effectively represent the discriminative features to ease the differences between kinship image pairs. Further, existing kinship methods only focus on a single resolution, and ignore the variability of resolutions in practical scenarios. To address these issues, we propose a multi-level dictionary pair learning (MLDPL) method to learn dictionary pairs by incorporating multiple resolution images for kinship verification. We learn dictionary pairs jointly by transforming discriminative features of image pairs into different coding coefficients in the same space, thereby reducing the differences between them. Further, multiple resolution images are incorporated into dictionary pair learning to effectively deal with resolution variations in kinship verification. Extensive experiments are performed on different kinship datasets to validate the efficacy of proposed MLDPL method. Experimental results show that MLDPL achieves competitive performance on all kinship datasets.