TL;DR
This paper proposes a novel face recognition system that improves accuracy and generalization by combining Embedded Hidden Markov Models (E-HMMs) with a selective ensemble strategy. Instead of using a single E-HMM, the system intelligently selects and combines multiple E-HMMs, each trained with different parameters, to create a more robust and adaptable face recognition model.
What Problem Does This Paper Address?
Automatic face recognition is a critical component of modern information security, offering a non-intrusive way to authenticate identities. While various methods exist, achieving high accuracy and, crucially, strong generalization ability (the ability to recognize new, unseen faces) remains a challenge. Traditional approaches often rely on a single model, which can be sensitive to variations in pose, lighting, and expression. The paper specifically targets the problem of model selection within the context of Embedded Hidden Markov Models (E-HMMs), a powerful technique for pattern recognition that can capture temporal and spatial information in images. The core issue is that E-HMMs, when applied to face recognition, can yield different performance based on parameter choices, and selecting the "best" model is not straightforward.
Methods
The paper introduces a two-part approach:
- Training Module:
- E-HMMs: The foundation is the Embedded Hidden Markov Model (E-HMM). An E-HMM models a face as a series of "super states" (e.g., forehead, eyes, nose) arranged vertically, with each super state containing "embedded states" arranged horizontally. Observation vectors (derived from 2D-DCT coefficients of image windows) are generated from these states. The model is trained using a 2D extended segmental K-means algorithm.
- Selective Ensemble: Instead of training just one E-HMM, the system trains a large number ($n$) of E-HMMs with diverse parameter settings. This creates a pool of models that capture different aspects of face features.
- Ensemble Module:
- Model Selection: From the pool of $n$ trained E-HMMs, a smaller subset ($m$) of the most accurate and diverse models is selected. This selection process prioritizes models that not only have high individual recognition rates but also complement each other by correcting errors made by other selected models. A weighted score is used, giving more importance to error correction ability.
- Ensemble Recognition: The selected $m$ models are then combined to recognize a probe image. The likelihood of the probe image belonging to each person is calculated using each of the $m$ models. These likelihoods are then combined using a weighted average, where weights can be set uniformly or, more effectively, based on the individual recognition rate of each model.
Datasets
The experiments were conducted using the ORL (Olivetti Research Ltd) face database. This database contains 400 gray-scale images of 40 individuals, with 10 images per person, at a resolution of 92x112 pixels. The authors describe several sample sets (A, B, C, D, E) derived from this database for training and testing, ensuring that different sets of gallery and probe images were used to evaluate the robustness of their method.
Key Insights and Ideas
- Diversity is Key: The paper strongly advocates for the idea that a diverse set of models, even if individually not the absolute best, can outperform a single, highly tuned model when combined. This is the core of the "selective ensemble" strategy.
- Error Correction is Crucial: The selection process explicitly considers a model's ability to correct errors made by other models in the ensemble. This suggests that models that capture complementary information are more valuable than those that simply replicate existing strengths.
- E-HMMs Aren't Universally Applicable: The results show that a single, commonly used E-HMM configuration (Mr) is not always optimal and can perform significantly worse than other models or the proposed ensemble. This highlights the need for a more sophisticated approach to model selection.
- Unequal Weighting Outperforms Average Weighting: When combining models, assigning weights based on individual model performance (recognition rate and error correction) yields better results than simple uniform averaging.
Key Results
- The proposed selective ensemble E-HMM algorithm achieved higher recognition rates compared to traditional single E-HMM based algorithms across all tested sample sets.
- For example, on sample set C, the ensemble achieved 99.17% recognition, while the best single model (Mc) achieved 98.5%. On sample set D, the ensemble reached 98.5% compared to MD's 99.5% (though the ensemble's average across sets was higher).
- The ensemble method also demonstrated lower variance in recognition rates across different sample sets, indicating better stability and generalization ability.
- When compared to other popular face recognition methods like Eigenfaces, Fisherfaces, Laplacianfaces, and CNNs, the proposed E-HMM ensemble achieved a superior average recognition rate (99.17% vs. 96.2% for CNN).
Limitations and Future Work
- Computational Complexity: The paper acknowledges that ensembling multiple E-HMMs leads to higher computational complexity. However, they suggest that the algorithm is "parallel virtually," implying that it can be efficiently implemented using parallel computing to mitigate this issue.
- Model Selection Sensitivity: While the selective ensemble improves generalization, the paper doesn't explicitly detail how to optimally set the weights ($w_1, w_2$) for selecting models or how to determine the ideal number of models ($m$) to include in the ensemble.
- Future Work: The authors suggest that further improvements could be made by exploring more sophisticated ensemble strategies and potentially by optimizing the parallel implementation for greater efficiency.
Summary updated