Discrete Laplacian Operator and Its Applications in Signal Processing

Waseem Waheed, Guang Deng, Bo LiuIEEE Access 2020 · DOI

Abstract

Fractional calculus has increased in popularity in recent years, as the number of its applications in different fields has increased. Compared to the traditional operations in calculus (integration and differentiation) which are uniquely defined, the fractional-order operators have numerous definitions. Furthermore, a consensus on the most suitable definition for a given task is yet to be reached. Fractional operators are defined as continuous operators and their implementation requires a discretization step. In this article, we propose a discrete fractional Laplacian as a matrix operator. The proposed operator is real (non-complex) which makes it computationally efficient. The construction of the proposed fractional Laplacian utilizes the DCT transform avoiding the complexity associated with the discretization step which is typical in the constructions based on signal processing. We demonstrate the utility of the proposed operator on a number of data modeling and image processing tasks.

Introduction

In signal and image processing, the Laplacian operator is a fundamental tool for tasks ranging from edge detection to smoothing. Standard integer-order derivatives are "local" operators, meaning they only consider immediate neighbors. Fractional calculus offers a way to create "non-local" operators that consider a wider neighborhood, effectively providing a "long memory" that can better capture complex patterns in noisy data.

However, most fractional operators are defined in continuous space, requiring a complex discretization step for use on digital data. This often involves solving filter design problems or dealing with complex numbers in the Fourier domain. This paper introduces a computationally efficient, discrete fractional Laplacian operator ($L_\alpha$) that leverages the Discrete Cosine Transform (DCT) to sidestep these issues.

Methods

The proposed operator is a matrix-based generalization of the standard 1D discrete Laplacian $L$. The standard Laplacian matrix is diagonalizable by the Discrete Cosine Transform (DCT-II), such that $L = M^T E M$, where $M$ is the DCT matrix and $E$ is a diagonal matrix of eigenvalues. The fractional version is created by raising these eigenvalues to the power of $\alpha$: $$L_\alpha = M^T E^\alpha M$$

Because the DCT uses real numbers, the resulting operator is computationally efficient and naturally handles "mirror-image" boundary conditions, which are preferred in image processing to avoid edge artifacts.

The fractional parameter $\alpha$ tunes the operator's 'memory' effect; lower values result in longer-range dependencies.

As shown in the figure above, the fractional order $\alpha$ acts as a tuning parameter. As $\alpha$ decreases, the operator's "memory" increases, allowing it to look further away from the target pixel to distinguish signal trends from high-frequency noise. This "tunability" allows the operator to be optimized for specific noise levels in a signal.

Applications and Results

The operator was evaluated on several standard benchmarks, including synthetic 1D signals and classic 2D grayscale images like Cameraman and Peppers, across various noise levels ($\sigma \in \{15, 25, 50\}$).

  • Trend Filtering: The fractional operator consistently achieved lower Mean Squared Error (MSE) than the standard Laplacian, particularly as noise levels increased. For example, on the "Peppers" image at high noise ($\sigma=50$), the fractional $\ell_1$ filter achieved an MSE of 25.06 compared to 29.50 for the standard version.
  • Image Sharpening: At $\alpha=0.5$, the operator produced visibly sharper results with fewer artifacts compared to standard methods.
  • Edge Detection: The fractional Marr-Hildreth edge detector proved significantly more robust to noise than its integer-order counterpart, maintaining clean edge lines where the standard version produced false positives.
  • Shock Filtering: In image enhancement, the fractional version provided better object segmentation and sharper boundaries.

Conclusion

The discrete fractional Laplacian provides a versatile, real-valued building block for signal processing. By treating the fractional order as a hyperparameter, the operator can be tuned to specific noise conditions. While the current implementation demonstrates strong performance in direct applications, its greatest potential lies in its integration as a regularizer within more complex variational frameworks or deep learning architectures for inverse problems.

Summary updated

BibTeX
@article{waheed2020discrete,
  author = {Waseem Waheed and Guang Deng and Bo Liu},
  title = {{Discrete Laplacian Operator and Its Applications in Signal Processing}},
  journal = {IEEE Access},
  year = {2020},
  doi = {10.1109/access.2020.2993577},
}
Figure