【摘要】 针对船舶受限空间烟雾检测中半透明目标特征提取不足与动态干扰鲁棒性差的问题，提出一种基于双路径注意力融合的自适应特征聚合网络（AFAN）。该方法以改进的Faster R-CNN为基线，首先通过通道-空间注意力模块（CSAM）增强关键语义特征，其次引入可移动方向注意模块（MAM）捕捉烟雾扩散方向特性，最后采用自适应权重融合策略整合双路径特征。实验表明，AFAN在船舶舱室烟雾数据集上达到90.42%mAP，较基线提升1.9%，在遮挡场景下误报率优于SSD和YOLOv5。该网络有效提升了烟雾特征的判别性与鲁棒性，为工业受限空间危险检测提供了解决方案。更多还原

【Abstract】 A dual-path attention fusion adaptive feature aggregation network(AFAN) is proposed to address the issues of insufficient feature extraction for semi-transparent targets and poor robustness to dynamic interference in ship confined space smoke detection. Based on an improved Faster R-CNN baseline, the method first enhances key semantic features through a channel-spatial attention module(CSAM), then captures smoke diffusion direction characteristics using a movable attention module(MAM), and finally integrates dual-path features via an adaptive weight fusion strategy. Experiments show that AFAN achieves 90.42% mAP on a s hip cabin smoke dataset, outperforming the baseline by 1.9%, and reduces false alarms in occluded scenarios compared to SSD and YOLOv5. The network effectively improves the discriminability and robustness of smoke features, providing a solution for hazardous detection in industrial confined spaces.更多还原