CORE-Det: Computation Reallocation for Large-Scene SAR Target Detection

Abstract

Large-scene synthetic aperture radar (SAR) target detection suffers from inefficient computation allocation because most processing is spent on background regions rather than target regions. We propose CORE-Det, a computation reallocation framework that dynamically redistributes processing across spatial regions, feature tokens, and training samples through a three-level sparsity-aware design tailored to SAR electromagnetic scattering. At the spatial level, the computation allocation via spatial distribution (COSD) module reconstructs a dynamic global spatial prior from sparse full-scene observations, performs patch-level screening before local detection, and fuses global and local confidence estimates to suppress background-generated target-like clutter. At the feature level, the computation allocation via feature value (COFV) module introduces SAR-aware mixture-of-experts routing, where specialist experts focus on target-dominant responses produced by strong scattering mechanisms, identity experts preserve structured contextual information at low cost, and zero experts suppress speckle-dominant and invalid activations. At the sample level, the decoupled error partitioning for adaptive retraining (DEPART) module mines hard cases caused by SAR scattering ambiguity and feeds module-specific error signals back to COSD and COFV, improving their ability to separate target-bearing regions from clutter and to distinguish target-dominant, contextual, and noise-dominant responses. Experiments on the self-built FUSAR-Airport and FUSAR-Harbor datasets with six baseline detectors show that CORE-Det consistently improves the accuracy-efficiency tradeoff. Compared with the corresponding baselines, it reduces GFLOPs by 88.92% and inference time by 51.97% on average, while improving mAP50 and recall by 9.59 and 8.94 percentage points (pp), respectively, and reducing false alarms by 19.61 pp.