Abstract: The authors gratefully acknowledge the support provided by SEP-DGEST for this research under grant DGEST 512.07-P.
Introduction
Despite decades of research, achieving 100% accuracy in face recognition remains a "holy grail," particularly in "non-cooperative" environments. These real-world scenarios require systems to contend with variations in Pose, Illumination, Expression, and Rotation (PIER). Most existing literature follows a purely engineering approach, treating face recognition as a matrix operation problem while ignoring the psychological and neuroscientific mechanisms that make human vision so efficient. Specifically, while standard systems struggle with low resolution or "fuzzy" data, human infants excel at recognizing faces using very limited visual information. This paper proposes a novel face recognition system (FRS) that bridges the gap between engineering and biology by mimicking how human newborns perceive faces.
Methodology: The Hough-KLT Approach
The authors introduce a hybrid approach called Hough-KLT combined with an unsupervised neural network.
- Face Feature Lines (FFL): Inspired by studies showing that newborns perceive a "diffuse world" and rely on basic curves and lines, the authors use the Hough Transform (HT) to extract the four most prominent lines in a low-resolution face image. This captures the configural relationship of the face, leaning into the "Margaret Thatcher Illusion" principle where humans process faces holistically rather than as isolated parts.
- Dimensionality Reduction: These line coordinates are concatenated with the original image data. The resulting high-dimensional vector is then compressed using Principal Component Analysis (PCA) and the Karhunen-Loeve Transformation (KLT) to keep only the most statistically significant features.
- Unsupervised Classifier: Instead of standard supervised backpropagation, the authors use a Self-Organizing Map (SOM). The SOM clusters the data into a 2D grid, which is then sharpened using a k-means algorithm. This unsupervised approach allows the system to "discover" the structure of face data, which proved more effective for generalization than forcing it into pre-defined labels.
Experimental Datasets
The study evaluates the proposed method against four classic benchmarks in the field, each representing different challenges of the PIER problem:
- ORL (Olivetti Research Laboratory): Contains variations in pose, illumination, and expression.
- YALE: Known for images with/without glasses and significant lighting changes.
- AR: Includes facial expressions, illumination changes, and occlusions like sunglasses or scarves.
- MIT: Features variations in pose and lighting.
Key Results and Discussion
The proposed SOM-Hough-KLT method showed impressive results, particularly on the ORL dataset, achieving 100% accuracy during the design phase and 90% accuracy during testing.
In head-to-head feature tests on the ORL set, the Hough-KLT features (89.43%) consistently outperformed standard Eigenfaces (85.12%) and Wavelet-based features (approx. 86%). Regarding classifiers, the SOM approach reached a peak performance of 92.86% on the ORL set, outperforming Euclidean Distance, Nearest Feature Line (NFL), and standard Feedforward Neural Networks (FFBP).
A primary limitation was the performance drop on the YALE database, which fell to 70% in testing. The extreme illumination changes in YALE proved difficult for the Hough-KLT feature extractor, as harsh lighting can "wash out" the distinct lines the system relies on.
Conclusion
By focusing on the "diffuse lines" used by newborns, the proposed system becomes less sensitive to high-frequency noise and minor pixel-level changes. Future work will integrate more comprehensive results from neuroscience to further refine how computer vision systems can emulate the robust, low-resolution recognition capabilities of the human brain, particularly to address extreme illumination challenges.
Summary updated