Introduction
High-dimensional feature vectors, such as Gabor filters or Local Binary Patterns (LBP), are essential for achieving high accuracy in face recognition. However, these "over-complete" features present significant challenges for practical deployment due to high memory requirements and slow matching speeds. Traditional hashing techniques often focus on distance preservation rather than classification accuracy. This paper introduces Bayesian Hashing, a framework that compresses high-dimensional features into compact binary codes by directly minimizing the Bayes error.
Methodology
The proposed system integrates three core components to optimize both efficiency and accuracy:
1. Bayesian Hashing
Unlike standard hashing methods, Bayesian Hashing seeks hash functions that minimize the theoretical minimum probability of misclassification, known as the Bayes error. The process involves a decorrelation projection using the Fisher criterion followed by a line-search to determine optimal bit-assignment thresholds.
2. Boosted FERNs
To classify the resulting binary codes efficiently, the authors employ Random FERNs—a variation of Naive Bayes. These are "boosted" using the GentleBoost algorithm, transforming the classification task into a series of simple Look-Up Table (LUT) summations. This approach is highly optimized for hardware, as it replaces complex matrix multiplications with memory reads.
3. Bit Permutation and Model Selection
To capture complex spatial relationships across the face, bits from different image patches are randomly shuffled. A Sequential Forward Floating Search (SFFS) algorithm then selects the most effective combinations of these shuffled models, ensuring high accuracy without redundant computation.
Experimental Results
The framework was evaluated on the FRGC v2 (Experiment 4) and LFW benchmarks.
Performance and Compression
On the FRGC dataset, the method achieved a 93.20% True Positive Rate (TPR) at a 0.1% False Positive Rate (FPR). Remarkably, it maintained high accuracy even at a 64x compression rate (reducing features to 4,000 bytes).
Efficiency
In comparative testing, Bayesian Hashing demonstrated a 10x speedup over Product Quantization (PQ), a standard industry technique for vector compression. On the LFW dataset, the method outperformed contemporary deep learning approaches like ConvNet-RBM, highlighting the effectiveness of discriminative binary encoding.
Conclusion
Bayesian Hashing provides a robust solution to the "curse of dimensionality" in face recognition. By aligning the hashing process with the classification objective, the authors achieve significant compression and speed improvements without sacrificing accuracy. Future work will focus on integrating this hashing technique directly into deep learning architectures for end-to-end optimization.
Summary updated