摘要
为提升水面目标的检测性能,保障智能船舶的航行安全,基于YOLOv3 (you only look once)提出了一种面向精准目标定位的水面目标检测算法.首先,利用残差密集模块改进了YOLOv3的头部网络,让不同特征间能够进行跨越式的信息交互.其次,将头部网络中的最近邻上采样层替换为了反卷积层,使得网络在训练过程中能够更加自主地学习特征缩放.最后,将普通的学习率衰减策略和余弦退火策略相结合,进一步提升网络的训练效果.利用真实水域下的图像数据对提出的方法进行训练和测试,实验结果表明:提出的方法将水面目标的检测精度提升了4.7%,实现了更加精准的目标定位.
To improve the detection performance of water surface objects and ensure the navigation safety of smart ships,a water surface object detection algorithm for accurate object location was proposed based on YOLOv3(you only look once).First,the head network of YOLOv3 was improved by using the residual dense module,which could enable comprehensive information fusion between different features.Then,the nearest neighbor upsampling layers in the head network was replaced with deconvolution layers,which could help the network learn feature scaling more autonomously during training.Finally,the ordinary learning rate decay strategy and cosine annealing strategy were combined to further improve the training effect of the network.The proposed method was trained and tested on the real water image data.Experimental results show that the proposed method improves the detection accuracy of water surface objects by 4.7%,achieving more accurate object location.
作者
冯辉
郭俊东
徐海祥
FENG Hui;GUO Jundong;XU Haixiang(Key Laboratory of High Performance Ship Technology of Ministry of Education,Wuhan University of Technology,Wuhan 430063,China;School of Naval Architecture,Ocean and Energy Power Engineering,Wuhan University of Technology,Wuhan 430063,China)
出处
《华中科技大学学报（自然科学版）》
EI
CAS
CSCD
北大核心
2023年第10期38-43,共6页
Journal of Huazhong University of Science and Technology(Natural Science Edition)
基金
国家自然科学基金资助项目(51979210,51879210)
国家重点研发计划课题资助项目(2019YFB1600600,2019YFB1600604)
中央高校基本科研业务费专项资金资助项目(2019III040,2019III132CG)。
关键词
智能船舶
目标检测
精准目标定位
残差密集模块
反卷积
smart ship
object detection
accurate object location
residual dense module
deconvolution
