Multi-Kernel Low-Rank Dictionary Pair Learning for Multiple Features Based Image Classification

Xiaoke Zhu, Xiao-Yuan Jing, Fei Wu, Di Wu, Li Cheng, Sen Li, Ruimin HuAAAI 2017 · OpenReview

Abstract

Dictionary learning (DL) is an effective feature learning technique and has led to interesting results in many classification tasks. Recently, by combining DL with multiple kernel learning, a few multi-kernel DL methods have been presented to solve the multiple feature representations based classification problem. However, how to improve the representation capability and discriminability of multi-kernel dictionary has not been well studied. In this paper, we propose a novel multi-kernel DL approach, named multi-kernel low-rank dictionary pair learning (MKLDPL). Specifically, MKLDPL jointly learns a kernel synthesis dictionary and a kernel analysis dictionary by exploiting the class label information. The learned synthesis and analysis dictionaries work together to implement the coding and reconstruction of samples in the kernel space. To enhance the discriminability of the learned multi-kernel dictionaries, MKLDPL imposes the low-rank regularization on the analysis dictionary, which can make samples from the same class have similar representations. We apply MKLDPL for multiple features based image classification task. Experimental results demonstrate the effectiveness of the proposed approach.

Introduction

In computer vision, a single image can be described by various features such as color histograms, textures (e.g., HOG), or shapes. While Dictionary Learning (DL) is a popular method for identifying underlying patterns in these features, traditional methods often face three major challenges:

  1. Non-linearity: Real-world image data is often non-linear, yet many models assume linear separability.
  2. Feature Integration: Effectively combining multiple feature types (e.g., color and texture) remains difficult.
  3. Incomplete Representation: Most methods focus solely on a synthesis dictionary for data reconstruction, neglecting the analysis dictionary necessary for efficient, discriminative coding.

MKLDPL addresses these gaps by creating a unified framework that handles non-linear, multi-feature data while ensuring that the resulting digital signatures for each class are consistent and discriminative.

Methodology

The MKLDPL framework is built upon three core components designed to maximize classification performance in non-linear spaces:

  • Multi-Kernel Learning (MKL): Instead of relying on a single kernel function, the model learns an optimal weighted combination of multiple kernels. This fuses different feature representations into a single, powerful non-linear space.
  • Dictionary Pair Learning (DPL): MKLDPL jointly learns a Synthesis Dictionary (to reconstruct samples in kernel space) and an Analysis Dictionary (to transform samples into coding coefficients for classification).
  • Low-Rank Regularization: By imposing a low-rank constraint (using the nuclear norm) on the analysis dictionary, the model forces samples from the same class to produce highly similar coding coefficients, significantly enhancing robustness.
The MKLDPL framework flowchart, illustrating the joint learning of weighted kernel matrices and dictionary pairs.

The optimization process utilizes an alternating strategy, employing the Alternating Direction Method of Multipliers (ADMM) for low-rank sub-problems and Quadratic Programming to update kernel weights.

Experimental Results

The model was evaluated on three standard benchmarks: Oxford Flowers17, Oxford Flowers102, and Caltech101. MKLDPL consistently outperformed state-of-the-art linear and multi-kernel methods across all datasets.

Comparison of classification accuracy (%) between MKLDPL and state-of-the-art methods across three benchmark datasets.

Key Findings

  • Superior Accuracy: On Flowers17, MKLDPL achieved 91.69% accuracy, a 3.5% improvement over the nearest baseline (DMKDL).
  • Noise Robustness: In tests with 30% random noise on Caltech101, MKLDPL maintained 72.13% accuracy, whereas other methods dropped as low as 46%.
  • Ablation Validation: Removing the low-rank constraint (the "MKDPL" variant) led to a consistent performance drop, confirming that low-rank regularization is essential for discriminative power.

Conclusion and Future Work

MKLDPL demonstrates that joint dictionary pair learning combined with low-rank constraints and multi-kernel fusion significantly improves image classification. While the $O(N^3)$ computational complexity may pose challenges for massive datasets, the model's robustness and accuracy make it a powerful tool for complex feature sets. Future research may extend this approach to video analysis and cross-modal retrieval tasks.

Summary updated

Abstract

Dictionary learning (DL) is an effective feature learning technique, and has led to interesting results in many classification tasks. Recently, by combining DL with multiple kernel learning (which is a crucial and effective technique for combining different feature representation information), a few multi-kernel DL methods have been presented to solve the multiple feature representations based classification problem. However, how to improve the representation capability and discriminability of multi-kernel dictionary has not been well studied. In this paper, we propose a novel multi-kernel DL approach, named multi-kernel low-rank dictionary pair learning (MKLDPL). Specifically, MKLDPL jointly learns a kernel synthesis dictionary and a kernel analysis dictionary by exploiting the class label information. The learned synthesis and analysis dictionaries work together to implement the coding and reconstruction of samples in the kernel space. To enhance the discriminability of the learned multi-kernel dictionaries, MKLDPL imposes the low-rank regularization on the analysis dictionary, which can make samples from the same class have similar representations. We apply MKLDPL for multiple features based image classification task. Experimental results demonstrate the effectiveness of the proposed approach.
BibTeX
@inproceedings{zhu2017multi,
  author = {Xiaoke Zhu and Xiao-Yuan Jing and Fei Wu and Di Wu and Li Cheng and Sen Li and Ruimin Hu},
  title = {{Multi-Kernel Low-Rank Dictionary Pair Learning for Multiple Features Based Image Classification}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2017},
  url = {https://openreview.net/forum?id=B1bhJxbObB},
}
Figure