Towards Frequency-Adaptive Learning for SAR Despeckling

TL;DR

The paper addresses despeckling in SAR imagery by tackling dataset bias between homogeneous and heterogeneous regions. It introduces SAR-FAH, a frequency-adaptive architecture that uses Haar wavelet decomposition to split content into low- and high-frequency sub-bands, with a neural ODE-based LFSP-ODE module for smooth low-frequency denoising and an asymmetric U-Net HFDE module for high-frequency refinement, connected via a Dynamic Attentive State-Space Fusion. The approach leverages NODE for structural fidelity, deformable convolutions for flexible edge handling, and VMamba-based global-context modeling to preserve textures, achieving superior performance on synthetic and real SAR data across standard quality indices and no-reference metrics. The work demonstrates that explicit frequency-domain processing combined with dynamic, cross-band fusion yields robust despeckling with preserved edges and textures, enabling more reliable downstream analyses in SAR applications.

Abstract

Synthetic Aperture Radar (SAR) images are inherently corrupted by speckle noise, limiting their utility in high-precision applications. While deep learning methods have shown promise in SAR despeckling, most methods employ a single unified network to process the entire image, failing to account for the distinct speckle statistics associated with different spatial physical characteristics. It often leads to artifacts, blurred edges, and texture distortion. To address these issues, we propose SAR-FAH, a frequency-adaptive heterogeneous despeckling model based on a divide-and-conquer architecture. First, wavelet decomposition is used to separate the image into frequency sub-bands carrying different intrinsic characteristics. Inspired by their differing noise characteristics, we design specialized sub-networks for different frequency components. The tailored approach leverages statistical variations across frequencies, improving edge and texture preservation while suppressing noise. Specifically, for the low-frequency part, denoising is formulated as a continuous dynamic system via neural ordinary differential equations, ensuring structural fidelity and sufficient smoothness that prevents artifacts. For high-frequency sub-bands rich in edges and textures, we introduce an enhanced U-Net with deformable convolutions for noise suppression and enhanced features. Extensive experiments on synthetic and real SAR images validate the superior performance of the proposed model in noise removal and structural preservation.

Figures (14)

• Figure 1: The illustration of restored results obtained by state-of-the-art methods using different training dataset. The first row is trained on a synthetic SAR dataset, while the second row is trained on a texture dataset. It demonstrates that increasing texture richness in the dataset helps with texture preservation.
• Figure 2: The framework of the proposed SAR-FAH model.
• Figure 3: Energy analysis of gradient-dominated image (first row) and texture-rich image (second row). It illustrates the homogeneous bias in dataset. (a) shows the original image. The energy distribution are shown in (b) for homogeneous regions, and (c) for heterogeneous regions.
• Figure 4: Synthetic texture images disrupted by speckle noise ($L=1$) to study statistical property after wavelet decomposition. (a) Clean SAR image and its decomposition by 2D Haar wavelet transform. (b) - (g) show the speckle noise ((b)), noisy image ((c)), low-frequency sub-band $X^\text{nLL}$ ((d)), low-frequency sub-band $X^\text{nLH}$ ((e)), $X^\text{nHL}$ ((f)), $X^\text{nHH}$ ((g)), with their histogram fitting by $\mathcal{GM}$ and $\mathcal{GGD}$ distribution, respectively. We also label the PSNR, SSIM value, and fitting parameters.
• Figure 5: Low-frequency structural preservation-ODE (LFSP-ODE) module.

Theorems & Definitions (2)

• Theorem 1
• proof