TL;DR
This paper introduces a novel face recognition system that leverages an "ordinal correlation" principle, which focuses on the relative ordering of pixel intensities rather than their absolute values. This approach demonstrates superior robustness to noise and outperforms traditional methods like Eigenfaces and 2DPCA, particularly in challenging scenarios.
What Problem Does This Paper Address?
Face recognition is a critical area in computer vision with applications ranging from security to personal identification. While existing methods like Eigenfaces and 2DPCA have achieved a certain level of maturity, they often rely on pixel intensity values directly. This makes them susceptible to noise (e.g., impulsive noise, bit errors) and certain non-linear transformations, leading to reduced accuracy. The paper aims to address this by developing a face recognition system that is inherently more robust to such variations by focusing on the relative ordering of pixel values.
Methods
The core of the proposed system is the "ordinal correlation principle." Instead of using raw pixel values, the method constructs "slices" that represent whether a pixel's value is less than another pixel's value. These slices are then aggregated into "metaslices" for specific regions of the image. The similarity between two images is then computed by comparing these metaslices using a "metadifference" operation, followed by a final aggregation step. This ordinal approach is designed to be insensitive to outliers and non-linear transformations.
To handle the computational cost of this method, images are downsized using bilinear interpolation. A key innovation is an automatic threshold selection algorithm that leverages intra-class (similarity between images of the same person) and inter-class (dissimilarity between images of different people) information from a training dataset. This algorithm aims to find an optimal threshold that balances true positives and false positives, improving recognition performance.
Datasets
The experiments were conducted on six datasets derived from the Yale database. These datasets were structured into training and testing subsets, with varying numbers of training images per individual and total training/testing images. Dataset 4, in particular, was used for evaluating the system's performance on noisy images.
Key Insights and Ideas
- Ordinality is Robust: Focusing on the relative order of pixel intensities, rather than their absolute values, provides a significant advantage in handling noise and variations in lighting.
- Region-Based Ordinality: Incorporating spatial information by dividing images into regions and applying ordinal measures within these regions enhances the discriminative power of the features.
- Automatic Thresholding is Crucial: The proposed automatic threshold selection method is vital for achieving optimal performance, as it dynamically adapts to the dataset's characteristics, unlike fixed thresholds.
- Ordinal Features Improve with More Data: The performance of the ordinal approach shows a notable improvement with more training data, suggesting that its features become richer and more discriminative with increased sample size.
Key Results
- The proposed ordinal correlation-based face recognition system outperformed Eigenface and 2DPCA approaches significantly in terms of precision and recall across various datasets.
- In Scenario 1 (fixed thresholds), the ordinal approach achieved higher precision and recall rates compared to Eigenface and 2DPCA. For example, on dataset 4, the ordinal approach achieved 98.2% precision and 96.3% recall, while Eigenface achieved 81.6% and 84.2%, and 2DPCA achieved 85.4% and 87.8%.
- The automatic threshold selection algorithm demonstrated stable performance across different subsets of the datasets, indicating its reliability.
- The system showed superior robustness to noise (salt and pepper, Gaussian, Poisson) compared to Eigenface and 2DPCA. Even with noisy testing images, the ordinal approach maintained high recognition rates.
Limitations and Future Work
- Computational Cost: A primary limitation acknowledged is the computational cost associated with generating slices for each pixel. The paper addresses this by downsampling images, but it remains a factor.
- Region Size Selection: The optimal size of the divided regions for comparison is not explicitly determined and is suggested as an area for future work, potentially using cross-validation algorithms.
- Dithering and Noise Effects: The paper suggests further investigation into the detailed effects of dithering and noise on the proposed measure.
- Color Images and Wavelet Techniques: Future work could extend the approach to color images and explore the use of wavelet techniques on the feature matrices to reduce computation and evaluate performance.
- Overlapping Regions: The paper mentions that increasing the overlapping percentage of region areas (e.g., 75% overlap) might lead to better recognition rates, which is a direction for future exploration.
Summary updated