Uncorrelated Multi-View Discrimination Dictionary Learning for Recognition

Xiao-Yuan Jing, Ruimin Hu, Fei Wu, Xi-Lin Chen, Qian Liu, Yong-Fang YaoAAAI 2014 · DOI

Abstract

This paper presents Uncorrelated Multi-View Discrimination Dictionary Learning (UMD²L), a framework designed to learn discriminative dictionaries from multi-view data. By integrating a novel uncorrelated constraint, UMD²L minimizes redundancy between views while maximizing class-specific reconstruction accuracy. Experimental results on benchmark datasets like Multi-PIE and AR Face demonstrate that UMD²L achieves superior recognition performance compared to existing multi-view subspace and dictionary learning methods.

Introduction

In computer vision, "multi-view" data—information gathered from different perspectives, sensors, or feature extractors—is common. For example, a person’s face might be captured from five different angles, or an object might be described by its shape, color, and texture simultaneously. While Dictionary Learning (DL) provides a robust way to represent data via sparse atoms, extending it to multi-view scenarios introduces two major hurdles: ensuring the learned dictionaries are discriminative enough for classification and reducing inter-view redundancy. UMD²L addresses these by jointly learning dictionaries that are both class-discriminative and mutually uncorrelated.

Methodology

UMD²L builds upon the Fisher Discrimination Dictionary Learning (FDDL) framework but extends it to the multi-view regime with two key innovations:

  • Discriminative Fidelity Term: Instead of a generic dictionary, UMD²L learns class-specific sub-dictionaries. It forces the model to ensure that a sample from Class A is reconstructed very well by Class A's sub-dictionary but very poorly by sub-dictionaries of other classes. This makes the reconstruction error itself a powerful tool for classification.
  • The Uncorrelated Constraint: This is the "secret sauce." The authors designed a mathematical constraint that penalizes correlation between dictionaries learned from different views. This forces each view's dictionary to find unique, non-redundant features that the other views might have missed.

The optimization follows a two-level strategy: it iteratively updates the variables for one view while holding others fixed, alternating between solving for sparse coding coefficients and updating the dictionary atoms using quadratic programming.

Experimental Setup

The model was evaluated on four diverse public datasets: Multi-PIE (face images across 5 poses), AR Face (frontal faces with varying lighting and occlusions), COIL-20 (rotated objects), and MNIST (handwritten digits).

Sample multi-view face images from the Multi-PIE dataset used for evaluation.

Key Results and Insights

UMD²L consistently outperformed both multi-view subspace learning (like MCCA and MvDA) and existing multi-view DL methods (like SliM² and MSRC):

  • Higher Accuracy: UMD²L achieved a mean recognition rate of 95.10% across all datasets, improving upon the best baseline by at least 1.67%.
  • Lower Correlation: Quantitative analysis showed that UMD²L produced dictionaries with significantly lower inter-view correlation coefficients. For example, on the Multi-PIE dataset, UMD²L's correlation was nearly half that of the SMBR method.
  • Redundancy Reduction: The paper demonstrates that simply having more data views isn't enough. If the dictionaries for different views are highly correlated, the additional views add no value. UMD²L’s ability to "de-correlate" these views is what drives its performance gains.

Conclusion and Future Work

UMD²L demonstrates that reducing redundancy is as critical as increasing discriminability in multi-view learning. By tying dictionary atoms directly to class labels and enforcing uncorrelation, the framework transforms a reconstruction task into a highly effective classification task. While the current approach relies on predefined features (like Gabor or LBP), future work may involve integrating these constraints into end-to-end deep learning architectures where features are learned automatically.

Summary updated

BibTeX
@inproceedings{jing2014uncorrelated,
  author = {Xiao-Yuan Jing and Ruimin Hu and Fei Wu and Xi-Lin Chen and Qian Liu and Yong-Fang Yao},
  title = {{Uncorrelated Multi-View Discrimination Dictionary Learning for Recognition}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2014},
  doi = {10.1609/aaai.v28i1.9134},
}
Figure