一种基于YOLO的轻量型多尺度船舶检测算法OA
A Lightweight Multi-scale Ship Detection Algorithm Based on YOLO
针对船舶检测中暴露出的参数量大、模型复杂、小目标检测效果欠佳的问题,提出了一种基于结构重参数化的轻量型船舶检测算法:RA-YOLO.首先,引入基于Ghost改进的结构重参数化网络RepGhost降低模型复杂度,提高检测效率实现轻量化.然后,采用Adown多尺度特征融合方案将深层特征和浅层特征整合,增强对语义线索的理解能力,提高小目标的检测精度.最后,为了得到高质量的预测框,提高对目标船舶的定位精度,使用 EIoU 损失函数优化边框回归.改进后的 RA-YOLO召回率和mAP@50 分别达到了83.4%、91.1%,而模型参数总量和每秒浮点运算次数则减少了30.0%和20.7%,提高了在资源有限设备上部署的可行性.
To tackle the challenges of large parameter counts,complex models,and suboptimal small target detection in ship detection,a lightweight ship detection algorithm based on structural re-parameterization,called RA-YOLO,is proposed.Firstly,the RepGhost backbone network,an enhancement of Ghost that employs structural re-parameterization,is used to reduce model complexity and improve detection efficiency.Then,the Adown multi-scale feature fusion is employed to integrate deep and shallow features,thereby enhancing the ability to understand semantic cues and improving the detection accuracy of small targets.Finally,the EIoU loss function is used to optimize bounding box regression,resulting in high-quality anchor boxes and enhanced localization accuracy for target ships.The enhanced RA-YOLO demonstrates increased recall and mAP@50,achieving 83.4%and 91.1%,respectively,while reducing parameters and floating point operations per second(FLOPs)by 30.0%and 20.7%,thus improving its feasibility for deployment on resource-constrained devices.
张开发;王玫;神显豪;唐超尘;阚瑞祥
桂林理工大学计算机科学与工程学院,广西 桂林 541006桂林理工大学物理与电子信息工程学院,广西 桂林 541006桂林理工大学物理与电子信息工程学院,广西 桂林 541006桂林理工大学物理与电子信息工程学院,广西 桂林 541006桂林电子科技大学 信息与通信学院,广西 桂林 541004
信息技术与安全科学
目标检测船舶检测结构重参数化特征融合
object detectionship detectionstructural re-parameterizationfeature fusion
《电讯技术》 2026 (1)
47-54,8
国家自然科学基金资助项目(62071135)广西科技重大专项(桂科AA23062035)
评论