Abstract
This paper introduces a robust face recognition system that replaces dataset-specific "Eigenfaces" with a fixed set of "natural" basis functions derived from general image statistics. By combining these multiscale filters with a biologically inspired Topographic Sparse Distributed Memory (SDM), the authors achieve expression-invariant recognition and near real-time performance. The system demonstrates that local feature representations, rather than global templates, provide superior tolerance to facial expressions and occlusions.
Introduction
In the mid-1990s, Principal Component Analysis (PCA)—popularized as "Eigenfaces"—was the standard for face recognition. However, PCA-based systems suffered from significant limitations: they required computationally expensive retraining whenever new faces were added, and they were highly sensitive to changes in facial expression, lighting, and partial occlusions. Because Eigenfaces analyze the entire image globally, even small local changes can drastically alter the resulting feature vector.
Rao and Ballard proposed a more universal and robust representation. Their approach shifts the focus from global templates to local "iconic" features, leveraging the statistical properties of natural scenes to create a general-purpose visual front-end that does not require constant retraining.
Methods
The proposed architecture consists of two primary components: a multiscale feature extraction layer and a topographic associative memory.
Natural Basis Functions
Instead of deriving basis functions from a specific face database, the authors performed PCA on thousands of arbitrary patches from natural scenes. They discovered that the dominant eigenvectors resulting from this process closely resemble oriented derivative-of-Gaussian functions.
The system utilizes 9 of these steerable filters across 5 different scales. For any given point on a face, these filters generate a 45-dimensional "iconic" vector representing local texture. A face is then represented as a collection of these vectors sampled on a radial grid centered on the face, rather than a single monolithic image.
Topographic Sparse Distributed Memory (SDM)
To store and retrieve facial identities, the system uses a modified version of Kanerva’s Sparse Distributed Memory. SDM acts as an associative memory that maps high-dimensional feature vectors to identities, allowing for "fuzzy" matching. This enables the system to recognize a face even when the input vector deviates due to changes in expression.
The memory is organized topographically, with separate SDMs for different facial regions (e.g., nose, mouth). This organization prevents features from different locations from interfering with each other (crosstalk) and significantly accelerates the search process. The authors implemented this on a Datacube MV200 pipeline processor, framing distance calculations as image convolutions to achieve real-time indexing at 30 frames per second.
Results
The system was evaluated on a database of 140 images featuring 20 individuals with varying facial expressions. Key findings include:
- Accuracy: The system achieved a peak recognition rate of 93.3% using 33 sampling points per face.
- Expression Invariance: The SDM successfully interpolated and identified individuals even when expression changes reduced feature vector correlation to approximately 0.45.
- Efficiency: Feature extraction was completed in roughly 0.5 seconds, supporting near real-time operation.
- Distinctiveness: The natural feature vectors proved highly distinctive, with low cross-individual correlation (~0.43).
Discussion and Conclusion
The primary insight of this work is the universality of natural image statistics; the filters optimal for general scene representation are also highly effective for specialized tasks like face recognition. This suggests a general-purpose "front-end" in the human visual system.
While the current system handles 2D rotations and scaling effectively, it remains limited by 3D head rotations. Future work may involve a "view-based" approach to handle profile shots and the integration of color-opponent Gaussian filters to enhance robustness. Despite these limitations, the combination of natural basis functions and topographic SDM provides a powerful framework for robust, real-time object recognition.
Summary updated