Details

Autor(en) / Beteiligte

• Gao, Linchun
• Wang, Shoujun
• Chen, Songgui
• Hu, Yuanye

Titel

AF-FPN: an attention-guided enhanced feature pyramid network for breakwater armor layer unit segmentation

Ist Teil von

• Multimedia systems, 2024-02, Vol.30 (1), Article 18

Ort / Verlag

Berlin/Heidelberg: Springer Berlin Heidelberg

Erscheinungsjahr

2024

Quelle

Alma/SFX Local Collection

Beschreibungen/Notizen

• The armor layer unit of the breakwater is a commonly used structure in ocean engineering. They are usually densely arranged and often have complex occlusion and overlap. These factors make traditional segmentation methods difficult to meet the requirements of high precision and efficiency. Unlike existing approaches to generic target segmentation, we propose a specific target segmentation method characterized by segmenting small and dense instances with similar characteristics. Our approach called the attention-guided enhanced feature pyramid network (AE-FPN) for breakwater armor layer unit segmentation, consists of two major components. The first component is an attention-guided (AM) module. Comprising both a channel context attention module (CCAM) and a spatial context attention module (SCAM), the AM module uses contextual information to allow the model to learn information about the region containing the target. The second component is semantic feature enhancement (SFE), wherein pyramid-like structures are employed to enhance semantic information. To assess the performance of the AE-FPN, a task-specific breakwater armor unit dataset named SUD2022 was released. Without any bells and whistles, the proposed AE-FPN achieved 75.2% AP on this dataset, which represents a 9.1% improvement over that of the Mask R-CNN. We also performed ablation experiments on the Cifar10 and COCO datasets to verify the generalization of our designed module. Large-scale experimental results on SUD2022 and COCO datasets demonstrate that the AE-FPN not only achieves excellent performance on breakwater armor layer unit segmentation but can also improve the performance of generic object segmentation.