Contextual constraints based kernel discriminant analysis for face recognition

Xian Wu, Xiaojun Wu, Zhenhua Feng, Xiaoqi Sun2013 · DOI

TL;DR

This paper introduces a novel face recognition method called Contextual Constraints based Kernel Discriminant Analysis (CCKDA). By combining non-linear kernel PCA with a contextual constraint-based LDA, CCKDA effectively captures both the non-linear variations in face images and the spatial relationships between pixels, leading to improved recognition accuracy.

What Problem Does This Paper Address?

Face recognition is a challenging task due to variations in pose, illumination, and expression. Traditional methods often fall into two categories: linear (like PCA and LDA) and non-linear (like kernel methods). While linear methods are simpler, they struggle with the complex, non-linear nature of faces. Non-linear methods, particularly kernel-based ones, can better represent these variations. However, a key limitation of existing methods, both linear and non-linear, is their failure to explicitly consider the spatial relationships among pixels within an image. This spatial context is crucial for understanding facial features and improving recognition. The paper aims to bridge this gap by incorporating contextual information into a powerful non-linear framework.

Methods

The proposed method, CCKDA, is a two-step feature extraction process:

  1. Kernel PCA (KPCA): This step maps the original high-dimensional face image data into a non-linear feature space. KPCA is used to capture the non-linear variations in facial images, which linear methods cannot effectively handle.
  2. Contextual Constraints based Linear Discriminant Analysis (CCLDA): After the non-linear mapping by KPCA, the CCLDA algorithm is applied in this new feature space. CCLDA is an extension of LDA that incorporates a "contextual constraint." This constraint leverages the spatial relationships between pixels by assigning higher weights to neighboring pixels and lower weights to distant ones when calculating within-class and between-class scatter matrices. This allows the algorithm to better model the local structure of facial features.

The paper also mentions that a nearest neighbor classifier is used for the final recognition phase.

Datasets

The proposed CCKDA method was evaluated on three standard face recognition datasets:

  • ORL Face Database: Contains 400 images from 40 identities, with variations in pose, illumination, and expression.
  • Yale Face Database: Consists of 165 images from 15 individuals, featuring different poses, illumination conditions, and expressions.
  • XM2VTS Face Database: Includes 295 subjects captured over multiple sessions, exhibiting pose and illumination variations.

The paper notes that for XM2VTS, face regions were extracted and resized.

Key Insights and Ideas

  • The Power of Combining Non-linearity and Context: The core insight is that effectively recognizing faces requires both capturing non-linear variations (addressed by KPCA) and understanding the spatial arrangement of facial features (addressed by contextual constraints in CCLDA).
  • Contextual Information is Crucial: Explicitly modeling the spatial relationships between pixels, especially by emphasizing local neighborhood information, significantly boosts recognition performance, particularly in challenging datasets.
  • Kernel Methods Excel with Complex Data: The experimental results consistently show that kernel-based approaches (KDA and CCKDA) outperform linear methods (LDA and CCLDA) on these face datasets, highlighting their ability to handle the inherent complexity of human faces.
  • CCKDA's Robustness: The proposed CCKDA method demonstrates good stability, achieving high recognition rates even with a small number of training samples.

Key Results

The paper reports that CCKDA consistently outperforms LDA, KDA, and CCLDA across all three tested databases (ORL, Yale, and XM2VTS).

  • ORL Database: CCKDA achieved recognition rates over 85% with as few as 3 training samples, showing good stability.
  • Yale Database: Methods incorporating contextual constraints (CCLDA and CCKDA) performed significantly better than LDA and KDA, especially given the rich expression variations in this dataset.
  • XM2VTS Database: Kernel-based methods (KDA and CCKDA) showed much higher recognition rates than linear methods. CCKDA slightly outperformed KDA on this dataset.

The authors attribute the overall success of CCKDA to the synergistic effect of using contextual information and the kernel-based non-linear representation.

Limitations and Future Work

  • Sensitivity to Illumination: The paper notes that kernel-based methods, including CCKDA, can be sensitive to illumination variations, as observed in the comparison between Yale and XM2VTS results.
  • Assumption of Context Preservation: The CCKDA method assumes that contextual information is preserved from the original image space to the kernel-induced feature space. The validity of this assumption in all scenarios might be a limitation.

The paper does not explicitly detail future work directions, but the identified limitation regarding illumination sensitivity suggests potential avenues for future research, such as developing illumination-invariant contextual features or robust kernel methods.

Summary updated

Abstract

This paper proposes a two-step subspace learning framework by combining non-linear kernel PCA (KPCA) and with contextual constraints based linear discriminant analysis (CCLDA) for face recognition. The linear CCLDA approach does not consider the higher order non-linear information in facial images, whereas the wide face variations posed by some factors, such as viewpoint, illumination and expression, existing in nonlinear subspaces may lead to many difficulties in face recognition and classification problems. To counteract the above problem, we incorporate the contextual information into kernel discriminant analysis by using KPCA in a two-step process, which provides more useful information for face recognition and classification. Experimental results on three well-known face databases, ORL, Yale and XM2VTS, validate the effectiveness of the proposed method. (4 pages)
BibTeX
@misc{wu2013contextual,
  author = {Xian Wu and Xiaojun Wu and Zhenhua Feng and Xiaoqi Sun},
  title = {{Contextual constraints based kernel discriminant analysis for face recognition}},
  year = {2013},
  doi = {10.1049/ic.2013.0014},
}
Figure