Visual recognition by subspace approaches on LBP features
Date of Issue2015
School of Electrical and Electronic Engineering
Traditionally, subspace approaches are applied on the holistic features. Recently, local binary pattern (LBP) has become popular because it is robust to illumination variations and alignment error. In this thesis, we exploit the advantages of both. Firstly, we propose a fast and accurate subspace face/eye detector and build a complete and fully automated face verification system on mobile devices. Secondly, to improve the robustness to image noise, we propose a noise-resistant LBP (NRLBP) with an embedded error-correction mechanism. Thirdly, we derive a data-driven LBP through optimizing the LBP structure directly using Maximal-Conditional-Mutual-Information scheme, towards the objective of reducing the LBP feature dimensionality and deriving discriminative LBP structures. Fourthly, to better remove unreliable dimensions of LBP histogram, we propose a patch-dependent/independent learning-based LBP. Lastly, to handle non-Gaussian distribution of LBP features, we propose a Chi-squared transformation that enhances the performance gain of subspace approaches on LBP features.
DRNTU::Engineering::Electrical and electronic engineering::Computer hardware, software and systems