Age Estimation with Local Ternary Directional Patterns

Abstract

Local texture descriptors have gained significant momentum in pattern recognition community due to their robustness compared to holistic descriptors. Local ternary patterns and its variants use a static threshold to derive textural code in order to improve Local Binary Patterns robustness to noise. It is not easy to select an optimum threshold in local ternary patterns and its variants for all images in a dataset or all experimental datasets. Local directional patterns uses directional responses to encode image gradient. Apart from considering only k significant responses, local directional patterns does not include central pixel in determining image gradient. Disregarding central pixel and \(8-k\) responses could result in lose of significant discriminative information. In this paper, we propose local ternary directional patterns that combines local ternary patterns and local directional patterns in determining image gradient. In local ternary directional patterns, the threshold is determined by the neighboring pixels and both significant, less significant responses and central pixel are considered in calculating image gradient. Evaluation of local ternary directional patterns on FG-NET dataset shows its robustness in local texture description compared to local directional pattern and local ternary pattern.