摘要
声纳图像目标检测在水下救援和资源勘探中具有重要意义。传统的声纳目标检测技术存在智能化程度低、鲁棒性差、实时性差、识别精度低等问题。尽管许多基于卷积神经网络的目标检测算法在自然图像中取得了很大的成功。然而,对于水下声纳图像来说,海底混响噪声干扰、前景目标区域像素占比低、成像分辨率差等问题对实现准确的水下目标检测提出了相当大的挑战。为了解决这些问题,文章基于YOLOv5目标检测模型提出了一种新的声纳图像目标检测器。首先,在原有Backbone的基础上基于多头注意力机制引入C3MHSA模块和SE注意机制,提高模型的收敛性和提取目标形状和空间有效特征的能力。此外,在Backbone中加入RFB模块,提高网络在高感受野存在的情况下学习重要信息的能力。实验结果表明,改进后的Yolov5网络的mAP@0.5值为98.9%,较原始YOLOv5模型有了全面大幅提升,明显优于现有方法。
Sonar image target detection is of great significance in underwater rescue and resource exploration.Traditional sonar target detection techniques have problems such as low intelligence,poor robustness,poor real-time performance and low recognition accuracy.Many Convolutional Neural Network(CNN)based target detection algorithms have achieved great success in natural images.However,for underwater sonar images,problems such as undersea reverberant noise interference,low pixel share in foreground target regions and poor imaging resolution pose considerable challenges to achieving accurate underwater target detection.To address these problems,this paper proposes a new target detector for sonar images based on the YOLOv5 target detection model.First,the C3MHSA module and SE attention mechanism are introduced based on the original Backbone based on the Multi-Head Self-Attention(MHSA)mechanism to improve the convergence of the model and the ability to extract target shape and spatially effective features.In addition,the RFB module was added to Backbone to improve the network’s ability to learn important information in the presence of high receptive fields.The experimental results show that the mAP@0.5 value of the improved YOLOv5 network is 98.9%which is comprehensively improved compared with the original YOLOv5 model,and thus the method is superior to existing ones.
作者
陈启北
韩路军
陈慧
Chen Qibei;Han Lujun;Chen Hui(Guangxi Nanning Liuyao Pharmaceutical Co.,Ltd,Nanning,Guangxi 530000,China;Guangxi Liuyao Group Co.,Ltd,Liuzhou,Guangxi 530000,China)
出处
《邢台职业技术学院学报》
2023年第5期54-59,94,共7页
Journal of Xingtai Polytechnic College
基金
广西研究生教育创新计划项目--“基于人工智能和协同蜜罐集群的海洋信息系统智能诱导安全防御体系的研究”,项目编号:YCSW2022289。
