DMSA-Net: Dual-Domain Multiscale Attentive Network for Robust SAR Ship Detection

Yuxuan Liu, Xiaoli Ji, Hongguang Wei, Bingwen Qiu, Yuan Liu, Xiubao Sui, Qian Chen

Abstract: Deep learning has achieved remarkable progress in synthetic aperture radar (SAR) ship detection. However, critical challenges remain in mitigating noise interference and addressing multiscale object variations. To solve this, this article proposes a dual-domain multiscale attentive network for SAR ship detection (DMSA-Net). Our contributions are threefold. First, we design a spatial-frequency dual-stage feature extraction network, which innovatively constructs a dual-stage feature optimization architecture, with noise suppression achieved by a frequency-domain decoupling-reconstruction module in the first stage, and multiscale perception enhancement using a global-local enhancement module in the second stage. Second, we develop a wavelet decomposition-based multiscale dilated fusion (MSDF) pyramid, which embeds discrete wavelet transform sampling into bidirectional cross-scale interaction paths and introduces MSDF modules at critical hierarchical levels to enhance multiscale feature discrimination. Third, a shared-kernel-based dynamically weighted group normalization detection head is employed to significantly reduce computational costs and parameters while maintaining competitive accuracy. Extensive experiments on the high-resolution SAR image dataset (HRSID), the SAR ship detection dataset (SSDD), and SAR-Ship-Dataset datasets demonstrate the effectiveness of DMSA-Net, achieving AP50 values of 92.4%, 97.2%, and 94.9% respectively, outperforming existing methods in complex scenarios.

External IDs:doi:10.1109/taes.2025.3612933