Visual recognition by subspace approaches on LBP features

Abstract

Subspace approaches have been widely used in visual recognition. Traditionally, they are applied on holistic features derived by vectorizing raw-image pixels. However, the performance is often limited by rigorous image alignment and high computational cost. Recently, local binary pattern (LBP) becomes popular because of robustness to illumination variations and alignment error, and fast and easy feature extraction. However, it still has some limitations such as sensitivity to image noise, high feature dimensionality, unreliable estimation of LBP histogram and non-Euclidean feature space. Subspace approaches have been applied on LBP feature and partially resolved the problems. However, non-Gaussian distribution of LBP feature limits the performance gain brought by subspace approaches. In this thesis, we exploit the advantages of both subspace approaches and LBP feature. In Chapter 3, we enhance the performance of subspace approaches on holistic features, i.e. we propose a fast and accurate subspace face/eye detector. Then, we boost the performance of LBP feature by a noise-resistant LBP with an embedded error-correction mechanism in Chapter 4, and LBP structure optimization in Chapter 5 that better captures image characteristic. Finally, we improve the performance of subspace approaches on LBP features by two learning-based LBP features in Chapter 6, and a Chi-squared transformation in Chapter 7 that tackles the non-Gaussian distribution of LBP feature. These contributions are discussed in detail as follow. Firstly, we improve the subspace approaches on holistic features. We propose a natural and non-intrusive way to secure mobile devices, i.e. a complete and fully automated face verification iv Abstract system. The proposed subspace face/eye detector locates the eyes at a much higher precision than Adaboost face/eye detector at a comparable speed thanks to discriminative APCDA features and attentional cascade strategy. The proposed approach that determines the class-specific threshold without sacrificing training data for validation data further boosts the performance. The proposed system is systematically evaluated on O2FN, AR and CAS-PEAL databases, and achieves a better recognition performance at a comparable speed compared with other approaches. Secondly, we improve the noise robustness of LBP feature. LBP feature is sensitive to noise. Local ternary pattern (LTP) partially solves the problem. However, both LBP and LTP treat the corrupted patterns as they are. Thus, we propose a noise-resistant LBP to preserve the image local structures in presence of noise. We encode the small pixel difference as an uncertain bit first, and then assign its value to form possible uniform codes that represent image micro-structures. In addition, we find that some image patterns such as lines are not captured in uniform codes. We then propose an extended noise-resistant LBP to capture line patterns. Both proposed approaches are shown more resistant to noise than other LBP variants. Thirdly, we propose to optimize the LBP struct...