In this study, we propose an end-to-end detection model based on cascaded spatial priors and dynamic upsampling for ship detection tasks in day–night transition environments, named the Cascaded Group and Dynamic Upsample-DEtection TRansformer (CGDU-DETR). To address the limitations of traditional methods in complex lighting conditions (e.g., strong reflections, low light), we designed a novel CG-Net model based on cascaded group attention and introduced a dynamic feature upsampling algorithm, effectively enhancing the model’s ability to extract multi-scale features and detect targets in complex backgrounds. The experimental results show that the CGDU-DETR achieves an AP of 93.4% on the day–night transition dataset, representing a 2.86% improvement over YOLOv12, and a recall of 95.2%, representing a 24.44% improvement over YOLOv12. Particularly for complex categories such as cargo ships and law enforcement vessels, the CGDU-DETR significantly outperforms YOLOv12, with improvements of 35.9% in AP and 63.7% in recall. Moreover, generalization experiments on the WSODD public dataset further validate the robustness of the model, with the CGDU-DETR achieving an AP of 95.1%, representing an 11.6% improvement over YOLOv12. These results demonstrate that the CGDU-DETR has significant advantages in ship detection tasks under day–night transition environments, effectively handling complex lighting and background interference, and providing reliable technical support for all-weather maritime surveillance.

Vessel detection technology based on marine remote sensing imagery is of great importance. However, it often faces challenges, such as small vessel targets, cloud occlusion, insufficient data volume, and severely imbalanced class distribution in datasets. These issues result in conventional models failing to meet the accuracy requirements for practical applications. In this paper, we first construct a novel remote sensing vessel image dataset that includes various complex scenarios and enhance the data volume and diversity through data augmentation techniques. Secondly, we address the class imbalance between foreground (small vessels) and background in remote sensing imagery from two perspectives: the sensitivity of IoU metrics to small object localization errors and the innovative design of a cost-sensitive loss function. Specifically, at the dataset level, we select vessel targets appearing in the original dataset as templates and randomly copy–paste several instances onto arbitrary positions. This enriches the diversity of target samples per image and mitigates the impact of data imbalance on the detection task. At the algorithm level, we introduce the Normalized Wasserstein Distance (NWD) to compute the similarity between bounding boxes. This enhances the importance of small target information during training and strengthens the model’s cost-sensitive learning capabilities. Ablation studies reveal that detection performance is optimal when the weight assigned to the NWD metric in the model’s loss function matches the overall proportion of small objects in the dataset. Comparative experiments show that the proposed NWD-YOLO achieves Precision, Recall, and ${\mathrm{AP}}_{50}$ scores of 0.967, 0.958, and 0.971, respectively, meeting the accuracy requirements of real-world applications.

Ship wake rotation detection technology based on remote sensing images has a wide range of application value in ship supervision, trajectory prediction and maritime rescue. However, objects in natural scenes usually present continuous and smooth features, while the contours and textures of ship wakes appear as discrete lines. This makes the wake have strong edge (discontinuity of image feature distribution) features but lack texture information, so the general detection model is difficult to perceive the wake completely, and low-dimensional edge features may also be ignored in deep networks. In addition, the significant directionality of wake also puts forward higher requirements for the accuracy of angle prediction. To address these issues, an efficient wake rotation detection algorithm is proposed in this study. Firstly, we construct a discrete edge focused feature extraction module. This module features parallel edge branches that capture spatial features of the image while sensing changes in pixel gradient intensity, thereby obtaining a richer feature representation. Secondly, we construct an edge guided context attention fusion module. It uses the edge information in lowdimensional features as a guide to adaptively fuse contextual features, effectively enhancing the model's global perception capability. Finally, we design a diagonal constrained probabilistic IoU. It further enhances the description of rotation angle and shape by considering the relative diagonal distance of minimum external horizontal rectangle, especially when dealing with square like objects. Experimental results show that our algorithm achieves accurate and stable ship wake detection, and has excellent robustness in weak objects and complex backgrounds. The mean Average Precision of this method for optical remote sensing wake detection reaches 95.1%, and the average error of angle prediction is reduced from 15.35$^{\circ }$ to 7.16$^{\circ }$. In addition, the algorithm also achieves excellent performance in general remote sensing scenarios, which verifies its excellent adaptability and generalization.

Pinpointing elusive and minor target vessels from satellite-based images is recognized as a considerable obstacle in the specialized areas of computer vision and the examination of remote sensing imagery. The majority of existing methods are based on the YOLO architecture, which relies on manually designed anchor points and nonmaximum suppression (NMS) postprocessing. The detection of small targets in a single scene, the phenomenon of “catastrophic forgetting” due to the streaming of data, and the issue of an “information bottleneck” present significant challenges in this field. In order to address these issues, we propose the following solutions. A hybrid explicit spatial prior MH-Net network based on Manhattan distance is designed. By decomposing the self-attention matrix and the spatial attenuation matrix, the spatial correlations of different directions and positions are captured, thus effectively alleviating the problem of catastrophic forgetting. We propose an online convolutional reparameterization efficient layer aggregation networks cross-stage fusion network. Through equivalent transformations, the complex network architecture is compressed into a single linear layer. The network groups and processes input features in parallel, integrating both low-dimensional and high-dimensional features to alleviate the information bottleneck problem. The prediction head of the model uses the DINO decoder and applies contrastive denoising to remove useless prediction boxes. This allows the proposed MHOE-DETR model to avoid thresholding and NMS), reducing the model’s computational complexity. The experimental results demonstrate that the MHOE-DETR algorithm, designed for this purpose, markedly enhances the detection performance of small and indistinct targets in private remote sensing datasets. The average accuracy, recall, and AP50 reached 96.3%, 91.4%, and 95.4%, respectively, while maintaining a low GFLOPS value (54.4 G) and parametric count (77.3 M). These findings offer substantial technical justification for the implementation of sea area management and maritime safety monitoring strategies.