A Frequency-Gated Decoupled-Synergy Network for SAR target detection

The presence of inherent speckle noise in Synthetic Aperture Radar (SAR) imagery considerably complicates the detection of diverse targets, resulting in a pronounced degradation in the performance of conventional detection algorithms. To address these issues, this letter proposes a novel network FGDS-Net (Frequency-Gated Decoupled-Synergy Network) that employs a frequency-domain transformation to filter effective information from SAR images. By combining the local features generated by Convolutional Neural Networks (CNNs) with the global features from Vision Transformers (ViTs), our method effectively mitigates the impact of noise. Furthermore, through a process of down-sampling and subsequent up-sampling, we disentangle and then fuse different features, which aligns multi-scale features and enhances the model’s semantic and feature understanding capabilities. This align strategy improves the model’s ability to discriminate between various objects. Experimental results on the SARDet-100K dataset demonstrate that the proposed spatio-frequency encoding strategy enhances the model’s target representation capability in complex scenes, validating the effectiveness of the method.