Abstract
The 2D Quaternionic Fourier Transform (QFT), applied to a real 2D image, produces an invertible quaternionic spectrum. By conserving uniquely the first quadrant of this spectrum, it is possible, after inverse transformation, to obtain a 2D quaternion image that generalizes the classical 1D analytical image. From this quaternion image, we compute the corresponding correlation product and apply the direct QFT to obtain the 4D Wigner-Ville distribution (WVD) of this analytical signal. This approach yields a local quaternion Wigner-Ville distribution spectrum, providing a robust framework for space-frequency analysis in multi-dimensional imagery.
Introduction
In 1D signal processing, the "analytical signal" constructed via the Hilbert transform is a fundamental tool for extracting instantaneous amplitude and phase. However, generalizing this to 2D images or 3D hyperspectral data is notoriously difficult. Standard complex numbers, possessing only one imaginary unit ($i$), are insufficient to capture the directional dependencies and multi-dimensional symmetries inherent in higher-dimensional data.
This paper addresses the challenge of creating a robust space-frequency representation for images. While the classical Fourier Transform identifies global frequency content, it lacks local spatial context. The Wigner-Ville Distribution (WVD) is a powerful tool for local analysis, but its extension to higher dimensions requires a more sophisticated mathematical toolkit: hypercomplex numbers like quaternions and octonions.
Methodology
The proposed framework utilizes Quaternion Algebra, where a number consists of one real part and three imaginary parts ($i, j, k$). The process begins with the 2D Quaternionic Fourier Transform (QFT), a "two-sided" transform where the signal is processed between two exponential kernels. This transform is invertible and produces a full quaternionic spectrum.
Constructing the Analytical Image
The core innovation lies in the "quadrant trick." In 1D, an analytical signal is formed by suppressing negative frequencies. In 2D, this generalizes to suppressing all but one quadrant of the quaternionic frequency plane. By preserving only the first quadrant of the QFT spectrum and performing an inverse transform, the authors produce a 2D quaternionic analytical signal.
This signal is composed of the original image, two partial Hilbert transforms (capturing horizontal and vertical features), and one total Hilbert transform. Following this, a quaternionic correlation product ($\rho$) is defined, and the QFT is applied to it, resulting in a 4D distribution that maps the local quaternionic spectrum across the image.
Key Results and Insights
The authors successfully formulated the 4D Wigner-Ville distribution for quaternionic signals. A key insight is that the QFT obeys specific Hermitian symmetry rules, which are crucial for correctly reconstructing the analytical signal without information loss.
The paper visualizes how a single real image can be decomposed into four distinct quaternionic components ($f, i h_1, j h_2, k h$). These components represent different directional energy gradients of the original image, providing a much richer local description than standard complex-valued transforms.
Furthermore, the authors demonstrate that for 3D signals, such as hyperspectral cubes, the logic extends to Octonions (8-component hypercomplex numbers). This involves seven imaginary units to represent the various directional Hilbert transforms required for 3D analytical signal construction.
Conclusion and Future Work
The Quaternionic Wigner-Ville Distribution provides a powerful "local spectrum" that allows researchers to see how frequency content changes at specific coordinates $(x, y)$ in an image. While the mathematical framework is robust, the authors acknowledge significant computational complexity when moving to higher dimensions like octonions.
Future work will likely focus on algorithmic optimizations to enable real-time processing and applying this framework to specific computer vision tasks such as image segmentation and texture analysis in hyperspectral imagery, where directional spectral features are paramount.
Summary updated