PINN-Inspired Self-Supervised Framework for Speckle Reduction in SAR Images

초록

This paper proposes a self-supervised synthetic aperture radar (SAR) despeckling framework for heterogeneous scenes and limited-data conditions. The proposed method combines a structure-aware dual-TV term and a speckle-aware self-supervised term, and links them through indicator-driven adaptive weighting derived from the physical and statistical characteristics of speckle noise and structural terrain information. From this viewpoint, the framework is PINN-inspired, but not in the standard sense of solving an explicit physical forward partial differential equation. Instead, it adopts the PINN principle of embedding a governing residual-type constraint into training and extends it by modulating that constraint with pixel-wise indicators that reflect local structure–speckle properties. The proposed method was validated on both synthetic speckled SAR and real single-look satellite SAR data. In addition to ablation, noise-level robustness, and cross-validation analyses, representative mixed-terrain examples were examined from three complementary viewpoints: full-image restoration, speckle suppression in homogeneous regions, and structure preservation in densely textured regions. The results showed that the proposed framework provides a more favorable balance between speckle suppression and structural preservation than the compared baselines. © 2008-2012 IEEE.

키워드