Optimal Bayesian Hashing for Efficient Face Recognition

Qi Dai, Jianguo Li, Jun Wang, Yurong Chen, Yu-Gang Jiang

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.

The Bayes error region represents the theoretical minimum misclassification probability between two distributions, which Bayesian Hashing aims to minimize.

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).

ROC curves on the LFW dataset demonstrate that Bayesian Hashing-Boosted FERNs (SFFS-8) achieves state-of-the-art performance, outperforming several contemporary deep learning and high-dimensional feature methods.

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

Abstract

In practical applications, it is often observed that high-dimensional features can yield good performance, while being more costly in both computation and storage. In this paper, we propose a novel method called Bayesian Hashing to learn an optimal Hamming embedding of high-dimensional features, with a focus on the challenging application of face recognition. In particular, a boosted random FERNs classification model is designed to perform efficient face recognition, in which bit correlations are elaborately approximated with a random permutation technique. Without incurring additional storage cost, multiple random permutations are then employed to train a series of classifiers for achieving better discrimination power. In addition, we introduce a sequential forward floating search (SFFS) algorithm to perform model selection, resulting in further performance improvement. Extensive experimental evaluations and comparative studies clearly demonstrate that the proposed Bayesian Hashing approach outperforms other peer methods in both accuracy and speed. We achieve state-of-the-art results on well-known face recognition benchmarks using compact binary codes with significantly reduced computational overload and storage cost.
BibTeX
@inproceedings{dai2015optimal,
  author = {Qi Dai and Jianguo Li and Jun Wang and Yurong Chen and Yu-Gang Jiang},
  title = {{Optimal Bayesian Hashing for Efficient Face Recognition}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2015},
}
Figure