SAR Image Object Detection Based on Cascaded Denoising and Counterfactual Self-Distillation

Abstract

Background noise in synthetic aperture radar (SAR) images can severely degrade object detection performance. Although many methods employ complex image denoising strategies, target features in SAR imagery are tightly coupled with background noise. Meanwhile, excessive pixel-level smoothing inevitably weakens critical scattering details, resulting in the loss of discriminative and localization-relevant information. To address the above challenges, a cascaded denoising and counterfactual self-distillation framework for SAR object detection (CDC-DETR) is proposed. By jointly optimizing multiple components of the detection model, CDC-DETR builds a multistage denoising architecture that suppresses noise while preserving critical target details. Specifically, a frequency-domain feature fusion module is introduced, which adaptively integrates features from different frequency bands through a gated mechanism. In addition, a background-aware encoder denoising module is designed to compute feature statistics guided by target semantic priors, enabling weighted suppression of interference from different background regions. This design encourages the model to focus on target areas while preserving relevant contextual information. Furthermore, to enhance discriminative target feature extraction, CDC-DETR incorporates a counterfactual self-distillation mechanism that derives counterfactual information from deformable attention maps and guides self-correction during training, thereby preventing the loss of semantic cues in background–target coupled regions. Extensive experiments on MSAR and CETC38-SAR datasets demonstrate the efficiency of the proposed method when compared with state-of-the-art techniques.