PGSNet: Physics-Guided Scatter-Preserving Network for Large-Scale SAR Object Detection

Abstract

Target detection using Synthetic Aperture Radar (SAR) remains challenging: multiplicative speckle can obscure shallow features, while isolated scatterers may be overlooked under conventional fixed-step subsampling processing. Attention-based fusion methods and standard ImageNet-to-SAR transfer methods typically treat high-frequency clutter as salient structures. This paper proposes PGSNet, a single-stage detector that incorporates SAR physical prior knowledge across three complementary depths. The Physics-Guided Deparameterized Stem Network (PGF-Stem) connects the original intensity with a gradient-based ratio edge map, and aligns these two data streams via a 1×1 projection, without the need to learn a gating mechanism. Spatial-to-Depth Convolution (SPD-Conv) then reorganizes the weak scattering responses along the channel dimension, thereby preventing these responses from being spatially averaged due to resolution degradation. The coefficient-of-variation-guided feature calibration (CGFC) ultimately employs the classic speckle reliability index to reweight the feature pyramid map. Built upon the generalized focal loss (GFL) framework and trained using ResNet-50 on the SARDet-100K dataset across 12 training cycles, PGSNet achieved an mAP of 58.2% and an APS of 53.1% on the COCO dataset, outperforming DenoDet V2 (56.71%) and ConvNeXt-B with MSFA (56.4%). The detector still runs at 36.9 FPS on the RTX 3090.