Objective Synthetic aperture radar (SAR) has become an indispensable tool for marine surveillance due to its all-weather, day-and-night imaging capabilities. However, automatic ship detection in SAR imagery remains highly challenging, primarily because of strong coherent speckle noise, drastic variations in target scale, and complex background interference in inshore scenarios. Existing lightweight models often struggle to maintain high detection accuracy and robustness under these varying conditions. To address these challenges, this paper proposes an improved lightweight detection model—SSD-YOLO—build upon the YOLOv8n architecture, designed to improve detection performance in complex marine scenarios.

Method The SSD-YOLO model constructs a hierarchical "Reinforce-Purify-Fuse" synergistic framework by integrating three theoretically grounded innovative modules. First, to address the challenge of weak target responses in heavy noise, a parameter-free target response enhancement module (C2f_SimAM) is incorporated into the backbone network. Leveraging an energy-function-based SimAM attention mechanism inspired by neuroscience, this module adaptively calculates 3D attention weights, effectively enhancing the linear separability of target neurons while suppressing background clutter without introducing any additional model parameters. Second, a lightweight anti-noise feature aggregation module (SAR_SPPF) is deployed at the end of the backbone. Unlike traditional spatial pyramid pooling, this module employs Ghost Convolutions to improve computational efficiency and introduces a novel dual-path pooling strategy. Theoretical derivations based on the Gamma distribution properties of SAR speckle noise show that combining small-kernel average pooling (for smoothing background variance) with large-kernel max pooling (for retaining strong scattering peaks) maximizes the signal-to-noise ratio. Third, a direction-sensitive multi-scale fusion module (C2f_DSConv) is integrated into the neck network. This module replaces standard convolutions with direction-sensitive depthwise separable convolutions. By applying kernels oriented at 0°, 45°, and 90°, it effectively captures the anisotropic texture and geometric structures of ships, thereby significantly improving the model’s ability to detect densely distributed and multi-scale targets.

Results Extensive experiments were conducted on the SSDD and HRSID datasets. On the SSDD dataset, the model achieved 97.0% precision, 96.4% recall, 99.0% mAP@0.5, and 74.3% mAP@0.5:0.95. Despite these substantial performance gains, the model remains highly lightweight, maintaining approximately 3.01M parameters and 7.9G FLOPs. Generalization experiments on the more challenging, high-resolution HRSID dataset demonstrated superior robustness, yielding an mAP@0.5 of 94.9%. Quantitative analysis of the feature map SNR further validated the internal rationale behind the proposed synergistic framework.

Conclusion This study presents an efficient and robust solution for SAR ship detection by integrating lightweight anti-noise processing, target response enhancement, and direction-sensitive feature capture strategies.