At present, deep learning has been widely used in SAR ship target detection, but the accurate and real-time detection of multi-scale targets still faces tough challenges. CNN-based SAR ship detectors are challenged to meet real-time requirements because of a large number of parameters. In this paper, we propose a lightweight, single-stage SAR ship target detection model called YOLO-based lightweight multi-scale ship detector (LMSD-YOLO), with better multi-scale adaptation capabilities. The proposed LMSD-YOLO consists of depthwise separable convolution, batch normalization and activate or not (ACON) activation function (DBA) module, Mobilenet with stem block (S-Mobilenet) backbone module, depthwise adaptively spatial feature fusion (DSASFF) neck module and SCYLLA-IoU (SIoU) loss function. Firstly, the DBA module is proposed as a general lightweight convolution unit to construct the whole lightweight model. Secondly, the improved S-Mobilenet module is designed as the backbone feature extraction network to enhance feature extraction ability without adding additional calculations. Then, the DSASFF module is proposed to achieve adaptive fusion of multi-scale features with fewer parameters. Finally, the SIoU is used as the loss function to accelerate model convergence and improve detection accuracy. The effectiveness of the LMSD-YOLO is validated on the SSDD, HRSID and GFSDD datasets, respectively, and the experimental results show that our proposed model has a smaller model volume and higher detection accuracy, and can accurately detect multi-scale targets in more complex scenes. The model volume of LMSD-YOLO is only 7.6MB (52.77% of model size of YOLOv5s), the detection speed on the NVIDIA AGX Xavier development board reached 68.3 FPS (32.7 FPS higher than YOLOv5s detector), indicating that the LMSD-YOLO can be easily deployed to the mobile platform for real-time application.

Due to the existence of domain shift, the synthetic aperture radar (SAR) automatic target recognition (ATR) model trained on simulated data will have significant performance degradation when applied to real-world measured data. To bridge the domain gap, this paper proposes a single-domain generalization (SDG) method based on multilevel data augmentation (MLDA), enabling SAR-ATR models that have been fully trained on simulated data to be generalized to unseen real SAR data. The proposed method aims to enhance the model’s generalizable capability through three key components: (1) pixel-level augmentation, which enriches data distribution via random Gaussian noise injection in the spatial domain and high-frequency perturbation in the frequency domain to enhance pixel-level diversity; (2) feature-level style augmentation, which probabilistically mixes instance-wise feature statistics, generating hybrid-styled feature maps to enhance feature-level diversity; (3) domain-adversarial training, which constructs an adversarial framework between a feature extractor and discriminator to enforce the learning of domain-invariant representations. Experiments on two simulation-to-reality SAR datasets demonstrate that the proposed method outperforms existing baselines and other SDG algorithms, achieving state-of-the-art (SOTA) performance on both datasets (96.76% accuracy on the public SAMPLE dataset and 93.70% accuracy on the self-built S2M-5 dataset).

Open-set synthetic aperture radar (SAR) automatic target recognition must not only classify familiar targets accurately, but also reject target categories that are absent from training. This task is particularly challenging because aspect-dependent scattering produces substantial and often multi-modal intra-class variation, while the unknown categories available for validation cannot fully represent the open space encountered at deployment. To address these coupled difficulties, we propose the Dynamic Adaptive Multi-scale ASC Network (DAMS-ASCNet), a unified representation–distribution–decision framework for open-set SAR recognition. First, a dynamic adaptive multi-scale angle-sensitive convolution module learns sample-dependent directional and scale responses, enabling the feature extractor to preserve informative scattering variations while suppressing irrelevant disturbances. Second, multiple prototypes are introduced to model the aspect-induced submodes of each known class, providing a more faithful class representation than a single-center description. Third, open-set decisions are performed using class-conditional prototype similarities and class-conditional calibration, with rejection thresholds estimated exclusively from validation data. The revised evaluation separates the effects of representation, training objective, prototype modeling, and calibration, and reports threshold-dependent operating-point metrics separately from threshold-independent score-discrimination metrics. Experiments on the MSTAR open-set benchmark show that, on the original open-set split, the proposed framework with validation-only class-conditional calibration achieves $80.33\pm 2.96\%$ open-set accuracy and $81.27\pm 5.48\%$ unknown recall over three seeds; additional held-out class compositions reveal substantial sensitivity to the choice of unknown classes. Controlled ablations further show that the matched training objective materially affects the calibrated operating point, while the architecture-level advantage over an ordinary-convolution counterpart is modest; excessive feature compactness can also be detrimental to open-set separability.