A variety of local structure patterns and their hybridization for accurate eye detection

Abstract

In this paper, existing local structure patterns (LSPs) are reviewed and categorized into two types: intensity-based LSPs (I-LSPs) and gradient-based LSPs (G-LSPs). I-LSPs include local binary pattern (LBP), modified census transform (MCT) and generalized binary pattern (GBP) methods that compare the intensities of differently selected neighboring pixels within a 3×3 neighborhood with differently-formed reference intensities, encoding 256, 511, and 19,162 binary patterns respectively. G-LSPs include local gradient pattern (LGP), modified gradient pattern (MGP), and generalized gradient pattern (GGP) methods that compare the gradient magnitudes of differently selected neighboring pixels within a 3×3 neighborhood with differently-formed reference gradients, encoding 256, 511, and 19,162 binary patterns respectively. We extend all these LSPs to multi-scale block LSPs (MB-LSPs) that concatenate multiple block-based LSPs. Finally, we propose several hybrid LSPs that combine I-LSPs and G-LSPs by means of the AdaBoost feature selection. In experiments using AR564, BioID, ColorFERET and LFW databases to evaluate the eye detection accuracy of the proposed LSPs, the I-LSPs were good for detecting the eyes in low quality images, the G-LSPs were good for detecting the eyes in high quality images, and the hybrid LSPs achieve state-of-the-art eye detection accuracy across image qualities.