基于改进YOLOv5的船舶目标检测算法-《集美大学学报（自然科学版）》

文章信息/Info

作者:: 丘锐聪1; 2; 周海峰1; 2; 陈颖1; 2; 张兴杰3; 黄金满4; 翁卫征5; （1.集美大学轮机工程学院，福建厦门 361021；2.福建省船舶与海洋工程重点实验室(集美大学)，福建厦门 361021；3.集美大学航海学院，福建厦门 361021；4.厦门安麦信自动化科技有限公司，福建厦门 361021；5.厦门三丰鑫科技有限公司，福建厦门 361001）

Author(s):: QIU Ruicong1; 2; ZHOU Haifeng1; 2; CHEN Ying1; 2; ZHANG Xingjie3; HUANG Jinman4; WENG Weizheng5; （1.School of Marine Engineering,Jimei University,Xiamen 361021,China;2.Key Laboratory of Shipping and Ocean Engineering of Fujian Province(Jimei University),Xiamen 361021,China;3.School of Navigation,Jimei University,Xiamen 361021,China;4.Xiamen Anmaixin Automation Technology Co.,Ltd.,Xiamen 361021,China;5.Xiamen Sanfengxin Technology Co.,Ltd.,Xiamen 361001,China）

Keywords:: object detection of ship; YOLOv5s; deformable convolution; improved loss function

摘要:: 为了解决船舶目标因宽高比较大而导致在特征提取过程中容易丢失细节信息、提取较多无关信息，以及小型船舶在图像中区域占比较小导致检测困难等问题，本文提出ShipsYOLOv5船舶目标检测算法。首先，利用可变形卷积特性设计C3_DCN模块并引入主干特征提取网络，通过增加采样偏移量来适应船舶目标形状特点，抑制无关背景信息影响，提升网络特征提取能力；然后，引入基于Wasserstein距离的度量方法改进边界框损失函数，提升模型对小型船舶目标的检测性能；最后，用Seaships船舶数据集进行验证实验，结果表明，本文提出的ShipsYOLOv5算法相较于基准模型在精度、召回率和mAP@0.5上分别有1.9%、1.9%和2.3%的提升，且分别达到了98.3%、99.0%和99.1%，能满足船舶目标检测性能要求。

Abstract:: In order to solve the problems of losing detail/extracting irrelevant information caused by relatively high width to height ratios duning detection of terget ships,and the detection difficulties caused by small ships occupying a relatively small area in the image,this paper made improvements based on YOLOv5s and proposed ShipsYOLOv5 ship target prposses a detection algorithm based on shipsVOLOv5.First of all,a C3_DCN module is designed by using the deformable convolution feature and the key feature extraction network is introduced to adapt to the shape characteristics of the target ship and suppress the influence of irrelevant background information as well as to improve the network feature extraction capability,by adding sampling offset.Secondly,the loss function is improved by using a metric method based on Wasserstein distance to improve the bounding box loss function,thereby enhancing the model’s detection performance for small vessel targets.In this paper,Seaships dataset is used to verify the above improved method.The experimental results show that compared with the benchmark model,the accuracy,recall rate and mAP@0.5 of the proposed ShipsYOLOv5 algorithm have improved by 1.9%,1.9% and 2.3% respectively,reaching 98.3%,99.0% and 99.1%,respectively.It can meet the performance requirements of ship target detection .