面向无人水面艇的水下伪装目标检测方法

Underwater camouflage objects detection method for unmanned surface vessels

无人水面艇环境感知能力受限于待测目标背景复杂、形状多样和伪装等因素,常规方法难以在上述情况下准确对水下伪装目标准确检测与评价。针对探测场景的多样化和复杂化,基于多任务学习策略提出一种面向无人水面艇的轻量型伪装目标检测方法 MFLNet(Multi-Feature Learning Network),通过借助图像梯度感知任务来提升无人水面艇对水下伪装目标的检测能力。首先,将图像特征提取任务解耦为语义特征提取和梯度特征提取;然后,向高层语义特征引入图像梯度特征并通过多尺度通道注意力模块生成初始预测图;最后,经过逐层的特征修正生成对伪装目标的最终预测。实验结果表明:MFLNet在CAMO-Test和NC4K-Test数据集上,结构相似性度量Sα指标可达0.824和0.851,检测性能达到先进模型水平,相比同策略轻量化模型参数量减少65%,检测速度可达73.7帧/s,满足水下检测数据实时传送需求,具有一定的实际应用价值。

The environmental perception capability of unmanned surface vessels(USVs)is limited by factors such as the complexity of the background,diverse shapes,and camouflage of the targets to be detected.Conventional methods struggle to accurately detect and evaluate underwater camouflage objects in such scenarios.To address the diversification and complexity of detection scenarios,this paper proposes a lightweight camouflage objects detection method for unmanned surface vessels,called MFLNet(Multi-Feature Learning Network),based on a multi-task learning strategy.It enhances the UsV's ability to detect underwater camouflage objects by leveraging the image gradient perception task.Initially,the feature extraction task is decoupled into semantic feature extraction and gradient feature extraction.Then,image gradient features are introduced into high-level semantic features,and initial prediction maps are generated through the proposed Multi-Scale Context Attention module.Finally,accurate final predictions are generated through feature corrections at each layer.Experi-mental results show that MFLNet achieves Sα values of 0.824 and 0.851 on the CAMO-Test and NC4K-Test datasets,re-spectively.Compared to models with the same strategy but reduced parameter counts by 65%,MFLNet achieves a detection speed of 73.7 frames per second,meeting the real-time data transmission requirements for underwater detection and demon-strating practical application value.