Fa YOLO: fog-aware instance segmentation of ships with feature refinement in foggy scenes

Fa YOLO: fog-aware instance segmentation of ships with feature refinement in foggy scenes

연구 분야: Verification

학회: Multimedia Systems

Ship instance segmentation is essential for intelligent maritime navigation and traffic safety. However, under adverse weather conditions such as fog, the image quality from imaging devices degrades significantly, leading to poor performance of existing instance segmentation methods. To address this challenge, we propose FA YOLO, a ship instance segmentation framework based on interference suppression and feature refinement designed to enhance performance under foggy conditions. First, we propose a Multi-Scale Feature Aggregation Mamba (MFAM) module, which utilizes a state space modeling approach and a multi-scale channel aggregation gating mechanism to enhance long-range dependency modeling and global contextual representation. Second, we propose an Adaptive Fog Dehazing Module (AFDM), which utilizes parallel channel and spatial attention mechanisms along with a window-based multi-head self-attention strategy to suppress fog-related interference and improve focus on target regions. Third, we propose a Multi-Scale Perception-Guided Attention Module (MPAM), which integrates channel-position attention fusion, multi-window branch feature extraction and similarity measurement strategies to adaptively enhance and aggregate multi-scale contextual features. In addition, to address the lack of suitable foggy ship instance segmentation datasets in the community, we collected and annotated a new instance segmentation dataset of maritime ships under foggy conditions, FSISD. This dataset contains 10,249 ship images, covering common ship categories and environmental conditions. Experimental results on Foggy Cityscapes, FSISD and Foggy COCO-boat demonstrate that FA YOLO outperforms the baseline YOLOv8s in segmentation accuracy by 3.3%, 2.2% and 1.3%, respectively, confirming superior performance and strong generalization capability.