In traditional photography, noise is often additive: a random graininess layered on top of a clean image. But in coherent imaging systems—like Synthetic Aperture Radar (SAR), side-scan sonar, and ultrasound—noise behaves differently. It is multiplicative, meaning the intensity of the noise scales with the intensity of the signal. This phenomenon, known as "speckle," creates a granular texture that obscures fine details and complicates automated analysis.
Researchers across diverse fields have converged on a shared strategy to solve this: homomorphic processing. By mathematically transforming the nature of the noise, they turn a complex multiplicative problem into a standard additive one.
The Intuition: Linearizing the Noise
The core challenge of multiplicative noise is that it is signal-dependent. If $S$ is the clean signal and $N$ is the noise, the observed image $I$ is $I = S \times N$. Standard denoising filters, which are designed to "subtract" noise, struggle here because there is no constant noise component to subtract.
The breakthrough, as described by Xijian Fan, Qingwu Li, and He Huang in their work on side-scan sonar, is to apply a logarithmic transformation. Because of the identity $\log(S \times N) = \log(S) + \log(N)$, the log-domain version of the image treats the noise as a simple additive component.
This transformation "linearizes" the problem. Once the noise is additive, researchers can reach for a massive toolbox of existing algorithms—from classical Gaussian filters to advanced sparse representations—that were never originally intended for sonar or radar physics.
A Universal Pipeline
This homomorphic approach follows a consistent three-step pipeline across different imaging modalities:
- Logarithmic Mapping: Convert the multiplicative speckle into additive noise.
- Denoising: Apply a filter to the log-transformed data.
- Exponential Reconstruction: Apply an inverse transformation to return to the original intensity scale.
Liping Hu and Hongcheng Yi demonstrate this in Automatic Target Recognition, where they use the log-transform to prepare SAR images for feature extraction. Similarly, Jinli Suo and colleagues apply this to Optical Coherence Tomography (OCT), noting that this linearization allows them to use robust convex optimization techniques that would otherwise be mathematically intractable.
However, the process isn't a perfect mirror. Zengguo Sun and Rui Shi point out in their research that the log-transform introduces a statistical bias. To maintain image fidelity, a "mean adjustment" or bias correction must be applied during the reconstruction phase to ensure the denoised image matches the original's energy levels.
Bridging to Deep Learning
The most recent evolution of this idea integrates homomorphic transformations with Convolutional Neural Networks (CNNs). Standard CNNs are exceptionally good at removing additive Gaussian noise because their fundamental operation—convolution—is a linear process.
By using a homomorphic front-end, researchers like Giovanni Poggi and Luisa Verdoliva enable CNNs to "see" speckle as if it were standard additive noise. In SAR image despeckling through convolutional neural networks, they show that this allows the network to focus on learning structural details rather than struggling with the physics of multiplicative interference.
This synergy is also what allows Sanketh Vedula and Alexander Bronstein to push ultrasound imaging toward CT-scan quality. By aligning the noise characteristics with the optimization assumptions of deep learning, they achieve massive speedups in processing time while preserving the sharp edges necessary for medical diagnosis.
Beyond Modern Sensors
While the primary drivers of this research are high-tech sensors, the intuition is so robust it even applies to historical preservation. E. J. Wharton, Sos Agaian, and Karen Panetta utilized logarithmic image processing to restore 14th-century manuscripts. In their case, the "noise" isn't from a radar pulse but from the uneven fading and blurring of ancient ink. The homomorphic approach allows them to amplify faint text signals against noisy backgrounds, proving that the math of "un-multiplying" noise is a universal tool for clarity.