Side-scan sonar image despeckling based on block-matching and 3D filtering

Xijian Fan, Qingwu Li, He Huang, Min WangJournal of Image and Graphics 2012 · DOI

This paper, published in the Journal of Image and Graphics, presents a specialized approach to cleaning up side-scan sonar (SSS) images. By mathematically transforming complex acoustic noise into a more manageable form, the authors demonstrate a way to remove "speckle" while keeping the sharp details necessary for underwater exploration.

TL;DR

Side-scan sonar images are often obscured by "speckle" noise that follows a complex Rayleigh distribution. This paper introduces a pipeline that transforms this noise into a standard Gaussian form, allowing the powerful BM3D (Block-Matching and 3D filtering) algorithm to effectively denoise the images while preserving critical textures and edges.

What Problem Does This Paper Address?

Side-scan sonar is a workhorse for underwater mapping, wreck detection, and seafloor classification. However, because it relies on coherent acoustic waves reflecting off uneven surfaces, the resulting images are plagued by "speckle noise."

Unlike the additive noise common in digital photography, speckle is multiplicative—it scales with the signal intensity. Furthermore, it follows a Rayleigh distribution rather than a standard Gaussian (Normal) distribution. Traditional denoising filters often struggle with this; they either leave the image too grainy or blur out the very edges and textures (like the hull of a shipwreck or the ripples in sand) that researchers need to see.

Methods

The authors propose a multi-step pipeline to bridge the gap between complex sonar physics and state-of-the-art image processing:

  1. Statistical Transformation: Since the popular BM3D algorithm is designed for additive Gaussian noise, the authors first apply a power transformation (using a Gaussian smooth function) to convert the Rayleigh-distributed noise into a Gaussian distribution.
  2. Logarithmic Mapping: They then apply a log transform. This mathematically converts the multiplicative noise into additive noise, making the image compatible with standard denoising tools.
  3. Adaptive BM3D Filtering: This is the core of the process. It involves:
  4. Grouping: Searching the image for similar patches and stacking them into 3D blocks.
  5. Collaborative Filtering: Applying a 3D transform. Because the patches are similar, the "true" signal is sparse in this 3D space, while the noise is spread out. A hard threshold is applied to "kill" the noise.
  6. Aggregation: Reassembling the patches, using weighted averages where they overlap to reduce artifacts.
  7. Inverse Transformation: Finally, the authors reverse the log and power transforms to return the image to its original dynamic range.

Datasets

The study evaluates the algorithm using two types of data:

  • Simulated Data: An "analog" sonar image (showing a simple object and shadow) to which synthetic Rayleigh speckle noise was added at three intensity levels ($\sigma = 0.02, 0.04, 0.06$). This allowed for objective measurement against a "ground truth."
  • Real-World Data: Actual side-scan sonar imagery of a submarine shipwreck in a near-sea area, used to test the algorithm's performance on complex, authentic underwater structures.

Key Insights and Ideas

  • Mathematical Bridging: The "aha moment" is the realization that you don't need to reinvent denoising for sonar. By using specific power and log transforms, you can "trick" high-end optical denoising algorithms into working perfectly on acoustic data.
  • Exploiting Redundancy: BM3D is effective because images are repetitive. By looking at many similar patches simultaneously (the "3D" part), the algorithm can distinguish between a unique texture detail and a random noise spike.
  • Overcoming the "Gibbs" Effect: Many wavelet-based methods create "ringing" artifacts or "scratches" near sharp edges. The authors show that their collaborative 3D approach significantly reduces these visual errors, leading to much cleaner shadows and boundaries.

Key Results

  • Quantitative Superiority: In simulations, the proposed method achieved a Peak Signal-to-Noise Ratio (PSNR) of 39.77 dB (at $\sigma=0.02$), significantly higher than the Median filter (31.59 dB) and outperforming other advanced methods like SA-DCT (38.33 dB) and NLF (38.13 dB).
  • Qualitative Clarity: On the real-world shipwreck data, the method produced images where the hull structure and the "acoustic shadow" were much more distinct. Other methods either left a "mottled" grain or created a blurry "oil painting" effect that obscured fine details.

Limitations and Future Work

The authors acknowledge that real-world sonar noise is incredibly complex. While their model handles the primary speckle component well, they simplified the problem by ignoring certain residual noise sources.

Future Work: The authors suggest that further research is needed to analyze these residual noise components more deeply and to develop more automated ways to select the adaptive filtering parameters for different seafloor environments.

Summary updated

BibTeX
@article{fan2012side,
  author = {Xijian Fan and Qingwu Li and He Huang and Min Wang},
  title = {{Side-scan sonar image despeckling based on block-matching and 3D filtering}},
  journal = {Journal of Image and Graphics},
  year = {2012},
  doi = {10.11834/jig.20120110},
}
Figure