Cross-Layer Scattering-Aware Feature Fusion Network for SAR Target Detection

Abstract

Synthetic aperture radar (SAR) target detection is a fundamental task in intelligent SAR image interpretation and supports a wide range of remote sensing applications. However, existing methods remain limited in handling sparse and discontinuous scattering responses, substantial variations in target scale, speckle noise, low target contrast, and clutter-induced target-like responses. In this paper, we propose a cross-layer scattering-aware feature fusion network consisting of three components for SAR target detection. First, the selective deformable convolution deployment strategy applies modulated deformable convolutions only at selected backbone stage transitions, improving feature extraction for irregular scattering responses while avoiding extensive replacement of standard convolutions. Second, the cross-layer feature pyramid network is designed to preserve informative target cues during multiscale feature fusion. Its context-localization fusion module explicitly integrates high-level semantic context with low-level localization details to maintain discriminative information across feature levels. Third, the intensity-modulated head performs intensity-based negative-sample reweighting during training according to the high-intensity pixel ratio, improving classification discrimination in cluttered SAR scenes without introducing additional computation during inference. Extensive experiments on SAR-AIRcraft-1.0 and MSAR-1.0 demonstrate that the proposed method achieves superior performance over representative general object detectors and recent SAR-specific detectors. Ablation studies further validate the effectiveness of the proposed components.