Mixture of Bilateral-Projection Two-dimensional Probabilistic Principal Component Analysis

Fujiao Ju, Yanfeng Sun, Junbin Gao, Simeng Liu, Yongli Hu

Abstract

This paper introduces mixB2DPPCA, a novel probabilistic model that integrates mixture modeling with bilateral-projection two-dimensional Principal Component Analysis (2DPCA). By employing a mixture of local linear sub-models and projecting image data from both row and column perspectives, the algorithm effectively captures complex, non-linear variations in image datasets. This approach preserves spatial structure and mitigates the "curse of dimensionality" inherent in traditional vector-based methods. Experimental results on standard benchmarks demonstrate that mixB2DPPCA outperforms existing methods in both reconstruction accuracy and recognition performance.

Introduction

Dimensionality reduction is a fundamental task in computer vision and pattern recognition. While Principal Component Analysis (PCA) and its probabilistic variant (PPCA) are widely used, they face significant challenges when applied to real-world image data. First, PCA assumes global linearity, which is often violated by complex data manifolds, such as faces with varying expressions or lighting. Second, traditional methods require "vectorizing" 2D images into 1D vectors, which destroys spatial correlations and leads to high-dimensional spaces that are difficult to model efficiently.

Although 2DPCA addresses the vectorization issue by treating images as matrices, most existing versions are either unilateral (projecting from only one side) or lack a robust probabilistic framework. This paper proposes mixB2DPPCA to overcome these limitations by combining bilateral projection with a mixture-of-experts approach.

Methodology

The mixB2DPPCA model is built on several key innovations:

  1. Bilateral Projection: Unlike standard 2DPCA, mixB2DPPCA uses two loading matrices ($L$ for rows and $R$ for columns) to compress an image matrix $X$ into a smaller latent core $B$. This dual projection better preserves the intrinsic structural information of the image.
  2. Mixture Modeling: To account for non-linearity, the model represents the data manifold as a collection of $K$ local linear sub-models. Each image is assigned a probability of belonging to a specific sub-model, allowing the algorithm to approximate complex curved manifolds through "soft" clustering.
  3. Probabilistic Framework: By defining the model within a Bayesian context with Gaussian noise, it handles data uncertainty and noise more effectively than purely algebraic methods.
  4. Variational EM Algorithm: The authors derive a Variational Expectation-Maximization (EM) approach to iteratively estimate the model parameters, including loading matrices, mean matrices, and mixing proportions.

Experimental Evaluation

The model was evaluated on four public datasets: MNIST (handwritten digits), Yale Face Database, AR Face Database (including occlusions like sunglasses and scarves), and FERET Face Database. Performance was compared against Generalized Low Rank Approximation of Matrices (GLRAM) and a standard Mixture of PPCA.

Reconstruction Performance

A primary metric for evaluating dimensionality reduction is the average reconstruction error. As shown in the following figure, mixB2DPPCA consistently achieves lower error rates than competing methods across different numbers of mixture components ($K$).

Comparison of average reconstruction error across iterations for different mixture components (K=2, 5, 10).

The results indicate that as $K$ increases, the model's ability to capture local variations improves, leading to significantly better reconstructions compared to global models like GLRAM, which do not benefit from increased complexity in the same way.

Recognition and Visual Quality

Beyond quantitative metrics, the visual quality of the reconstructed images is a key differentiator. Reconstructions from mixB2DPPCA appear sharper and more faithful to the original features, especially in the presence of occlusions or varying expressions.

Visual comparison of original face images (top) and their reconstructions (bottom) using the proposed mixB2DPPCA model.

In recognition tasks using a Nearest Neighbor (1-NN) classifier on the latent cores, mixB2DPPCA achieved a recognition accuracy of 83.57% on the AR database ($K=10$), significantly higher than GLRAM (71.90%) and standard Mixture of PPCA (71.33%).

Conclusion

The mixB2DPPCA model successfully addresses the limitations of global linearity and image vectorization. By combining bilateral matrix projection with a probabilistic mixture model, it provides a powerful tool for image representation and recognition. Future work will focus on reducing computational complexity for high-resolution data and exploring extensions to higher-order tensors for video analysis.

Summary updated

BibTeX
@inproceedings{ju2016mixture,
  author = {Fujiao Ju and Yanfeng Sun and Junbin Gao and Simeng Liu and Yongli Hu},
  title = {{Mixture of Bilateral-Projection Two-dimensional Probabilistic Principal Component Analysis}},
  booktitle = {IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year = {2016},
  doi = {10.1109/cvpr.2016.483},
}
Figure