摘要
为了解决船舶图像目标检测任务中船舶目标尺度多样、背景环境复杂导致的漏检误检等问题,提出一种用于船舶目标检测的DSAMPN-YOLOv8n算法模型。引入动态蛇形卷积(dynamic snake convolution,DSC)和无参数注意力机制(simple parameter-free attention mechanism,SimAM),提高YOLOv8n网络在复杂环境下对检测目标的特征提取和融合能力;构建多尺度渐近特征金字塔网络(asymptotic feature pyramid network,AFPN)实现多尺度特征信息交换,增强特征融合效果,提高模型对不同尺度船舶目标的检测能力。使用Seaships船舶数据集对改进算法模型进行实验验证,结果表明,相较于YOLOv8n算法,优化后的算法DSAMPN-YOLOv8n的精确率、召回率和mAP@0.5分别提升了6.0%、6.7%、3.6%,实验结果验证了改进方法的准确性和有效性。
To address the challenges of missed and false detections in ship image target detection tasks caused by diverse ship target scales and complex background environments,a DSAMPN-YOLOv8n algorithm model for ship target detection was proposed.The model introduces dynamic snake convolution(DSC)and simple parameter-free attention mechanism to enhance the YOLOv8n network s ability to extract and integrate features in complex environments.Additionally,a multi-scale asymptotic feature pyramid network(AFPN)was constructed to facilitate multi-scale feature exchange,improving feature fusion and enhancing the model s capability to detect ship targets of different scales.The proposed algorithm model was experimentally validated using the Seaships dataset.Results showed that,in comparing with those from the YOLOv8n algorithm,the optimized DSAMPN-YOLOv8n algorithm achieves improvements in precision,recall,m AP@0.5 by 6.0%,6.7%,3.6%,respectively.These findings validate the accuracy and effectiveness of the proposed method.
作者
陈颖
周海峰
郑东强
张兴杰
黄金满
CHEN Ying;ZHOU Haifeng;ZHENG Dongqiang;ZHANG Xingjie;HUANG Jinman(School of Marine Engineering,Jimei University,Xiamen 361021,China;Key Laboratory of Shipping and Ocean Engineering of Fujian Province,Xiamen 361021,China;School of Marine Equipment and Mechanical Engineering,Xiamen 361021,China;School of Navigation,Jimei University,Xiamen 361021,China;Xiamen Anmaixin Automation Technology Co.,Ltd.,Xiamen 361001,China)
出处
《集美大学学报(自然科学版)》
2026年第2期176-188,共13页
Journal of Jimei University:Natural Science
基金
国家自然科学基金项目(51179074)
福建省自然科学基金项目(2021J01839)
集美大学安麦信产学研项目(S20127)。
