Multiple-Action Transform of Quaternion for Color Images

Rushi Lan, Xiaoqin Wang, Xiaolan Xie, Linfa LuIEEE Access 2019 · DOI

Introduction

In traditional digital image processing, color images are typically treated as a collection of three independent grayscale channels: Red, Green, and Blue. While computationally efficient, this approach fails to account for the intrinsic correlations and complex interactions between color channels. To address this, researchers have increasingly turned to Quaternionic Representation (QR), where each pixel is modeled as a four-dimensional number (a quaternion).

However, existing quaternionic operators—such as Clifford translation, rotation, and reflection—are often "shallow." They typically involve interactions between only two or three quaternions at a time, limiting their ability to capture the high-level features required for sophisticated computer vision tasks. This paper introduces the Multiple-Action Transform of Quaternion (MATQ), a novel mathematical framework that cascades quaternionic operators to extract deeper, more complex relationships within color data.

Methodology: The MATQ Framework

Inspired by the hierarchical structures of deep learning, such as Convolutional Neural Networks (CNNs), MATQ moves beyond single-operation transforms. It creates a multi-layered architecture by nesting basic quaternionic operators:

  1. Single Action Manifold of Quaternion (SAMQ): Operations involving two quaternions (e.g., Clifford translation).
  2. Double Action Manifold of Quaternion (DAMQ): Operations involving three quaternions (e.g., 3D or 4D rotation).

By cascading these actions, MATQ generates high-level descriptions of color channels, such as cross-differences and complex image contrasts. The framework is highly flexible, offering a significantly larger parameter space than standard fixed-parameter rotations.

The hierarchical architecture of the Multiple-Action Transform of Quaternion (MATQ), showing the cascading of functions to produce high-level quaternionic representations.

To validate the framework, the authors developed two specific pipelines: QDis_MATQ, a distance measure for impulse noise detection, and MATQ-BoW, a feature extraction method for image classification using a Bag-of-Words model.

Experimental Setup and Results

The MATQ framework was evaluated across several standard benchmarks for noise detection and texture classification. For impulse noise detection, the authors utilized classic images such as Lena, Peppers, and Fruits.

A processed quaternionic representation of the classic 'Lena' test image, used to evaluate the framework's performance in noise detection and feature extraction.

Key Performance Metrics

  • Impulse Noise Detection: The MATQ-based distance (QDis_MATQ) consistently outperformed standard modulus-based and rotation-based distances. At high noise levels (25%), MATQ achieved a significantly lower Number of Wrong Detections (NWD) compared to state-of-the-art quaternionic filters.
  • Texture Classification: The MATQ-BoW approach set new benchmarks across three major databases:
  • KTH-TIPS2-a: Achieved 70.6% accuracy (a 2.6% improvement over QMCBP).
  • USPTex: Reached 93.27% accuracy.
  • Outex TC00013: Reached 90.36% accuracy.

In all classification tasks, MATQ outperformed traditional Local Binary Patterns (LBP) and advanced quaternionic variants like QLBP and QLRBP.

Conclusion

MATQ represents a significant advancement in quaternionic image processing by shifting the focus from raw pixel intensities to the complex relationships between them. While the current architecture is manually designed, future research will explore learning strategies to automatically discover optimal MATQ structures. Additionally, the authors envision expanding MATQ to domains such as color image segmentation and RGB-D processing, where the four-dimensional nature of quaternions can effectively encode both color and spatial depth information.

Summary updated

BibTeX
@article{lan2019multiple,
  author = {Rushi Lan and Xiaoqin Wang and Xiaolan Xie and Linfa Lu},
  title = {{Multiple-Action Transform of Quaternion for Color Images}},
  journal = {IEEE Access},
  year = {2019},
  doi = {10.1109/access.2019.2938346},
}
Figure