摘要
针对目前SAR影像船舰目标检测方法存在多场景下检测精度不高、漏检、模型泛化能力差的问题,尝试以YOLOv8s网络为基础,提出新的注意力机制D-CBAM,并定义新的损失函数RIoU,以及将最新的可变形卷积DCNv4替换标准卷积,引入融合空间金字塔池化focal modulation networks来提升网络性能,提出的网络命名为DR_YOLOv8s++检测网络。为验证DR_YOLOv8s++网络的有效性和通用性,在SSDD、HRSID数据集上进行实验。结果表明,所提出算法的平均精度均值分别达到98%、97.5%,优于其他经典算法,模型性能提升明显,同其他目标检测算法相比,具有较强的泛化能力。
Aiming at the problems of low detection accuracy,missed detection and poor model generalization ability in multiple scenarios in the current detection methods for ship targets in SAR images,a new attention mechanism D-CBAM is proposed based on YOLOv8s network,a new loss function RIoU is defined,and standard convolution is replaced by the latest deformable convolution DCNv4.A focal modulation network based on spatial pyramid pool is introduced to improve network performance.The proposed network is named DR_YOLOv8s++detection network.In order to verify the effectiveness and universality of DR_YOLOv8s++network,experiments are carried out on SSDD and HRSID data sets.The results show that the average accuracy of the proposed algorithm reaches 98% and 97.5% respectively,which is superior to that of other classical algorithms.The performance of the model is significantly improved,and compared with other target detection algorithms,it has stronger generalization ability.
作者
杨明秋
陈国坤
董燕
左小清
YANG Mingqiu;CHEN Guokun;DONG Yan;ZUO Xiaoqing(Faculty of Land and Resources Engineering,Kunming University of Science and Technology,Kunming 650093,China;Key Laboratory of Plateau Remote Sensing,Yunnan Provincial Department of Education,Kunming 650093,China)
出处
《遥感信息》
北大核心
2025年第2期159-168,共10页
Remote Sensing Information
基金
云南省科技厅基础研究计划面上项目(202401AT070366)
云南省重大科技专项计划(202202AD080010)
国家自然科学基金(42161067)。
