TL;DR
This paper introduces a novel image descriptor called the Kernel Gabor-Based Weighted Region Covariance Matrix (KGWRCM) for face recognition. By weighting pixels based on their discriminative power and incorporating Gabor features within a kernelized framework, KGWRCM achieves state-of-the-art performance on benchmark face datasets, outperforming existing methods.
What Problem Does This Paper Address?
Traditional methods for face recognition often treat all pixels in an image region equally when extracting features. However, this assumption is flawed, as certain facial features (like eyes and mouth) are more critical for recognition than others (like cheeks or forehead). This paper addresses this limitation by proposing a weighted approach that assigns different importance to pixels, aiming to capture more discriminative information. Furthermore, to handle the complex, non-linear relationships present in face images, the authors extend this weighted approach into a kernelized framework, building upon the success of Gabor features in capturing rich visual properties.
Methods
The core of this work lies in the development of the Kernel Gabor-Based Weighted Region Covariance Matrix (KGWRCM). The process involves several key steps:
- Weighted Region Covariance Matrix (WRCM): Instead of a standard Region Covariance Matrix (RCM) where all pixels contribute equally, WRCM introduces a weighting matrix. This matrix is computed based on the similarity of each pixel within a face sample, effectively emphasizing more important pixels and deemphasizing less important ones.
- Gabor Features: To enhance the discriminative power of the features, Gabor wavelets are employed. These are known for their ability to capture spatial localization, orientation selectivity, and spatial frequency characteristics, which are crucial for visual recognition tasks. The Gabor features are obtained by convolving the image with a set of Gabor filters.
- Kernelization: To capture non-linear relationships and map features into a higher-dimensional space, a nonlinear kernel mapping is applied. This allows for a linear computation in the high-dimensional space, effectively handling complex patterns that linear methods might miss. The paper utilizes a Gaussian kernel.
- Distance Metric: A generalized eigenvalue-based distance metric is used to measure the dissimilarity between the computed covariance matrices.
The proposed KGWRCM method integrates these components: it partitions a face image into regions, extracts Gabor-based features, computes weighted region covariance matrices for these regions, and then kernelizes these matrices to compute distances for recognition.
Datasets
The proposed KGWRCM algorithm was evaluated on three standard face recognition datasets:
- ORL Face Database: Contains 400 images of 40 subjects, with 10 images per subject, exhibiting variations in pose and scale.
- Yale Face Database: Consists of 165 grayscale images of 15 individuals, with variations in expression and lighting.
- AR Face Database: A larger dataset with over 4,000 images of 126 individuals, featuring significant variations like illumination changes and occlusions (sunglasses, scarves).
The paper describes specific data splitting strategies for training and testing on each dataset, including resizing images and selecting random subsets for training and testing to ensure robust evaluation.
Key Insights and Ideas
- Pixel Importance Matters: The fundamental insight is that not all pixels contribute equally to face recognition. A weighted approach can significantly improve feature discriminability by focusing on salient facial regions.
- Gabor Features Enhance Discrimination: Combining Gabor features with region covariance matrices provides a richer representation that captures more essential visual information for distinguishing faces.
- Kernelization Unlocks Non-Linearity: For complex image data like faces, kernel methods are essential for capturing non-linear relationships and achieving higher accuracy, especially when combined with powerful feature descriptors.
- Weighted Kernelized Approach is Superior: The combination of weighted features, Gabor descriptors, and kernelization (KGWRCM) demonstrably outperforms its constituent parts and other established methods.
Key Results
The KGWRCM method achieved superior performance across all tested datasets:
- ORL Face Database: Achieved a mean recognition accuracy of 99.21%, outperforming KGRCM (98.41%) and other methods.
- Yale Face Database: Achieved a mean recognition accuracy of 79.20%, surpassing KGRCM (76.23%) and other approaches.
- AR Face Database: Achieved a mean recognition accuracy of 95.95%, which was 4.15% higher than KGRCM and significantly better than other methods.
The paper also highlights that the weighted approaches (WRCM and KGWRCM) generally outperform their non-weighted counterparts (RCM and KGRCM), underscoring the benefit of emphasizing important facial parts.
Limitations and Future Work
The authors acknowledge that, similar to KGRCM, the computational cost of KGWRCM is high due to the computation of high-dimensional matrices.
For future work, they propose developing an effective KGWRCM method with low computational complexity for face recognition. This suggests exploring more efficient ways to compute or represent the kernelized weighted region covariance matrices without sacrificing accuracy.
Summary updated