融合位置信息和上下文的水面目标检测方法  

Water Surface Object Detection Method that Combines Positional Information and Context

作　　者：马赛 解志斌[1,2] 邵长斌 MA Sai;XIE Zhibin;SHAO Changbin(Ocean College,Jiangsu University of Science and Technology,Zhenjiang 212003,China;Intelligent Marine Information Sensing and Transmission Laboratory,Zhenjiang 212003,China;School of Computer,Jiangsu University of Science and Technology,Zhenjiang 212003,China)

机构地区：[1]江苏科技大学海洋学院,江苏镇江212003 [2]智慧海洋信息感知与传输技术重点实验室,江苏镇江212003 [3]江苏科技大学计算机学院,江苏镇江212003

出　　处：《小型微型计算机系统》2024年第9期2221-2227,共7页Journal of Chinese Computer Systems

基　　金：国家自然科学基金项目(62276117)资助.

摘　　要：针对复杂多变的水面环境中,小目标检测精度低、漏检率高且检测平台计算资源有限的问题,提出了一种基于Efficientdet-D0融合位置信息和上下文的水面目标检测方法.首先,采用坐标注意力机制对主干特征提取网络的主要模块移动翻转瓶颈卷积进行改进,将目标的位置信息集成到通道注意力中,提高网络对水面小目标的检测能力;其次,在特征融合网络BiFPN中引入Cot模块,增强特征融合网络对特征图相邻和全局上下文的获取能力,进一步提高检测小目标的能力;最后,为优化预测网络训练,将预测网络激活函数替换为H-swish.在WSODD测试集中的实验结果表明,本文模型的mAP相比于原始模型提高了16.95%,漏检率下降明显,且本文模型参数量小于大多现有模型,证明了本文方法在水面目标检测模型中的有效性.Aiming at the problems of low detection accuracy,high missed detection rate and limited computing resources of detecti-on platforms in complex and changeable water surface environment,a water surface object detection method based on Efficientdet-D0 combines position information and context is proposed.Firstly,the coordinate attention mechanism is used to improve the main module of the backbone feature extraction network Efficientnet-B0 to mobile inverted bottleneck convolution,embed the object's p-osition information into the channel attention,improve the network's ability to detect small objects on the water surface;Secondly,t-he Cot module is introduced into the feature fusion network BiFPN,enhance the feature fusion network's ability to acquire feature map neighbors and global contexts,and further improve the ability to detect small objects;Finally,to optimize predictive network tr-aining,replace the predictive network activation function with H-swish.The experimental results in the WSODD test set show that the mAP of the proposed model is increased by 16.95%compared with the original model,the missed detection rate decreases signi-ficantly,and the number of parameters of the proposed model is smaller than that of most existing models,which proves the effecti-veness of the proposed method in the water surface object detection model.

关 键 词：水面目标检测 深度学习 Efficientdet-D0 位置信息 上下文信息 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程] TP391[自动化与计算机技术—控制科学与工程]