Answers are grounded in this paper and cite their sources.
Accurate identification of marine ship categories is of great significance for tasks such as maritime traffic control and criminal activity prevention. However, due to the small inter-class differences among marine ships and the potentially enormous intra-class differences under different scenarios, the model fails to correctly learn category discrimination features. This further affects the training efficiency of the detection network and reduces the model's classification performance. To address the problems of marine ship shape differences and marine ship size variations caused by imaging positions in high-resolution remote sensing images, we propose a marine ship detection algorithm based on adaptive feature extraction and multi-scale feature weighted fusion(SDAMNet). Considering the characteristics of marine ship targets in remote sensing images, the algorithm introduces an adaptive feature extraction module to adapt to the shape of marine ships, thereby enhancing the model's ability to extract features. Meanwhile, we also design a multi-scale feature weighted selection and fusion module, which uses high level features as weights for low-level features and combines information from different levels to obtain accurate location information and semantic information, enhancing feature representation. Experiments on Level 1 of the shipRSImageNet dataset demonstrate that the joint deployment of AFEM and MSFPN yields gains of 3.85%, 1.85%, and 1.88% in mAP50, mAP75, and mAP50:95 over the baseline, respectively. Ablation experiments on the HRSC2016 dataset corroborate this synergy: combining both modules achieves improvements of 1.66%, 3.32%, and 3.94% in mAP50, mAP75, and mAP50:95 over the baseline, with the pronounced gain at the stricter mAP75 threshold indicating enhanced localization precision, thereby verifying the generalization of the proposed method across different datasets.
No takes yet. Share an insight, caveat, or question.
Chen et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: