[Paper Review] Robust Face Recognition via Block Sparse Bayesian Learning

Taiyong Li, Zhilin Zhang|arXiv (Cornell University)|Jan 29, 2013

Face and Expression Recognition40 references4 citations

TL;DR

This paper proposes Block Sparse Bayesian Learning (BSBL) for robust face recognition by exploiting intra-block correlation in representation coefficients, significantly improving recognition accuracy and robustness over state-of-the-art sparse representation methods. Experiments on Extended Yale B, AR, and CMU PIE databases show BSBL achieves higher recognition rates, especially under occlusion and disguise, outperforming SRC and BSCO by up to 13.19% in challenging conditions.

ABSTRACT

Face recognition (FR) is an important task in pattern recognition and computer vision. Sparse representation (SR) has been demonstrated to be a powerful framework for FR. In general, an SR algorithm treats each face in a training dataset as a basis function, and tries to find a sparse representation of a test face under these basis functions. The sparse representation coefficients then provide a recognition hint. Early SR algorithms are based on a basic sparse model. Recently, it has been found that algorithms based on a block sparse model can achieve better recognition rates. Based on this model, in this study we use block sparse Bayesian learning (BSBL) to find a sparse representation of a test face for recognition. BSBL is a recently proposed framework, which has many advantages over existing block-sparse-model based algorithms. Experimental results on the Extended Yale B, the AR and the CMU PIE face databases show that using BSBL can achieve better recognition rates and higher robustness than state-of-the-art algorithms in most cases.

Motivation & Objective

• Address the limitation of existing sparse representation methods in handling intra-block correlation among training samples of the same subject.
• Improve face recognition robustness under challenging conditions such as occlusion, noise, and facial disguise.
• Leverage the block-sparse structure of face data to enhance representation accuracy and classification performance.
• Demonstrate that BSBL, a recently proposed Bayesian framework, outperforms existing block-sparse algorithms in recognition accuracy and stability.

Proposed method

• Formulate face recognition as a block-sparse representation problem, where training images of the same subject form a block in the dictionary matrix.
• Apply Block Sparse Bayesian Learning (BSBL) to estimate representation coefficients, explicitly modeling intra-block correlation among coefficients of the same subject.
• Use a hierarchical Bayesian model to learn the block structure and correlation patterns, enabling more accurate and robust sparse coding.
• Compute reconstruction error per subject using estimated coefficients and assign the test image to the class with minimum error.
• Integrate feature extraction (e.g., Gabor or PCA) before applying BSBL to reduce dimensionality while preserving discriminative information.
• Optimize the Bayesian model using variational inference to efficiently estimate hyperparameters and sparse coefficients.

Experimental results

Research questions

• RQ1 Can exploiting intra-block correlation in representation coefficients improve face recognition performance compared to standard block-sparse models?
• RQ2 How does BSBL compare to state-of-the-art sparse representation methods like SRC and BSCO under occlusion and noise?
• RQ3 Does BSBL maintain superior performance under real-world facial disguises such as sunglasses and scarves?
• RQ4 To what extent does the block-sparse Bayesian framework enhance robustness in low-dimensional or corrupted face image scenarios?

Key findings

• On the Extended Yale B database, BSBL achieved a recognition rate of 99.56% on $24 imes 21$-pixel images with 50% block occlusion, outperforming BSCO by 13.19%.
• Under 40% occlusion on $12 imes 10$-pixel images, BSBL achieved 45.28% recognition rate, compared to 30.33% for SRC and 20.88% for NN.
• On the AR database with scarf disguises, BSBL achieved 81.00% recognition at $27 imes 20$ resolution, significantly outperforming SRC (21.50%) and BSCO (44.00%).
• In the case of neutral expression with sunglasses, BSBL achieved 65.50% recognition at $42 imes 30$ resolution, surpassing NN (68.00%) and SRC (89.00%) at lower dimensions.
• BSBL consistently achieved the highest recognition rates across all tested conditions, including noise, occlusion, and disguise, demonstrating superior robustness.
• The recognition performance of BSBL remained stable and high even at low image resolutions (e.g., $6 imes 5$), indicating strong generalization capability.