摘要
针对海上目标姿态复杂且尺度多变,导致现有人工智能算法难以稳定检测的问题,提出一种融合深度监督与改进YOLOv8的海上目标检测算法.首先,设计了多尺度卷积模块,提取目标多种感受野的特征信息,减少漏检率;然后,添加深度监督网络,提高网络对深层类别信息及浅层位置信息的利用率,优化主干网络的目标特征提取性能;最后,在网络检测头部分引入通道注意力机制,过滤无关信息,增强对关键特征的识别率.在海上目标数据集中的实验结果表明,改进算法的mAP值达到93.69%,召回率达到85.16%,相比原模型分别提高了7.38、8.52个百分点,且优于对比的经典算法和新颖算法,检测时间约14 ms,满足海上实时目标检测需求,可为航运管理、预防海上事故等提供有效技术参考.
To address the unstable detection of marine targets challenged existing artificial intelligence algorithms due to the target s complex poses and variable scales,a detection approach based on deep supervision and improved YOLOv8 is proposed.A multi-scale convolution module is designed to extract the feature information of the target s multi-receptive fields and reduce the missed detection rate.Then,a deep supervision network is added to improve the utilization ratio of deep class information and shallow location information,thus optimizing the performance of the backbone network in target feature extraction.Finally,a channel attention mechanism is introduced into the detection head to filter the irrelevant information and enhance the recognition rate of key features.Experiments on the marine target dataset show that the mAP value and the recall rate of the proposed approach reach 93.69%and 85.16%,respectively,which are 7.38 and 8.52 percentage points higher than those of the original model,and the proposed approach outperforms both classical and novel algorithms.The detection time is about 14 ms,which meets the requirements of real-time marine target detection and provides technical support for shipping management and marine accident prevention.
作者
张建东
ZHANG Jiandong(Tianjin Communication Center,Northern Navigation Service Center,Maritime Safety Administration,Tianjin 300456,China)
出处
《南京信息工程大学学报》
CAS
北大核心
2024年第4期482-489,共8页
Journal of Nanjing University of Information Science & Technology
基金
交通运输部北海航海保障中心项目(2022-9)。
关键词
海上目标
深度学习
深度监督
多尺度卷积
通道注意力机制
marine targets
deep learning
deep supervision
multi-scale convolution
channel attention mechanism
