Despeckling

Despeckling is the process of reducing speckle noise in images produced by coherent imaging systems, such as synthetic aperture radar (SAR) and medical ultrasound. The goal is to suppress the granular noise while preserving important structural information and textures, thereby improving interpretability and downstream analysis. Despeckling is often treated as a noise reduction task that must balance noise suppression with edge preservation.

Speckle is typically modeled as multiplicative noise, where the observed image I is the product of the true reflectivity S and a speckle field N. This multiplicative nature distinguishes despeckling from additive noise reduction. In practice, techniques may operate in the original domain or after a transformation, such as applying a logarithm to convert multiplicative noise into an additive form, which can simplify filtering.

Despeckling methods span several families. Spatial-domain approaches include adaptive filters like the Lee, Kuan, and Frost filters, which use local statistics to tailor smoothing, and median-based methods that aim to reduce speckle while preserving edges. Transform-domain and model-based techniques involve homomorphic filtering (log transform followed by linear filtering and exponentiation), wavelet or curvelet thresholding, and variational or Bayesian methods. More recent advances employ nonlocal means, BM3D-type algorithms adapted for multiplicative noise, and deep learning models such as convolutional neural networks trained for despeckling. Hybrid and learning-based methods often yield superior edge preservation at higher noise reduction but require representative training data or careful parameterization.

Common evaluation metrics include the equivalent number of looks (ENL) for speckle suppression, as well as PSNR and SSIM for fidelity. Practical despeckling must balance noise reduction with preservation of edges and fine detail, and may introduce artifacts if over-applied. Applications span remote sensing, medical imaging, and underwater acoustics.