基于跨阶段双分支特征聚合的遥感小目标检测  

Remote Sensing Small Object Detection Based on Cross-stage Two-branch Feature Aggregation

作　　者：李杰 刘扬 李良 苏本淦 魏佳隆 周广达 石艳敏 赵振[1] Li Jie;Liu Yang;Li Liang;Su Bengan;Wei Jialong;Zhou Guangda;Shi Yanmin;Zhao Zhen(College of Information Science and Technology,Qingdao University of Science and Technology,Qingdao 266061,China;Qingdao Zichai Boyang Diesel Engine Company,Qingdao 266700,China;China Unicom Qingdao Branch,Qingdao 266001,China)

机构地区：[1]青岛科技大学信息科学技术学院,山东青岛266061 [2]青岛淄柴博洋柴油机股份有限公司,山东青岛266700 [3]中国联通青岛市分公司,山东青岛266001

出　　处：《系统仿真学报》2025年第4期1025-1040,共16页Journal of System Simulation

基　　金：国家自然科学基金(62201314,62201571);山东省自然科学基金(ZR2020QF007);强链计划(23-1-2-qdjh-18-gx)。

摘　　要：针对YOLOv8在遥感小目标检测中由目标尺度差异和复杂背景引起的漏检和误检问题,提出了基于跨阶段双分支特征聚合的遥感图像小目标检测方法。融合卷积算子中的全局共享权重与注意力中特定token的上下文感知权重,获得高频局部信息和低频全局信息;使用轻量级MLP捕获全局远程依赖关系,并设计并行跨阶段可学习视觉中心机制捕获输入图像的局部角区域信息;设计多维度残差注意力机制,聚合两个并行分支的输出特征,捕获像素级的成对关系以及跨通道和跨空间信息。实验结果表明:该模型在DIOR和RSOD数据集上的mAP分别达到了73.8%和98.1%,比对比方法分别提高了1.3%和2.1%。Aiming at YOLOv8's leakage and false detection problems caused by target scale difference and complex background in remote sensing small target detection,this paper proposes a remote sensing image small target detection method based on cross-stage two-branch feature aggregation.The global shared weights in the convolution operator and the context-aware weights of specific tokens in the attention are fused to obtain high-frequency local information and low-frequency global information;the global remote dependencies are captured using a lightweight MLP,and the parallel cross-stage learnable vision center mechanism is designed to capture the information of the local corner regions of the input image;a multidimensional residual attention mechanism is designed to aggregate the output features of two parallel branches to capture pixel-level pairwise relationships as well as cross-channel and cross-space information.The experimental results show that the proposed model achieves 73.8%and 98.1%mAP on DIOR and RSOD datasets respectively,which is 1.3%and 2.1%higher than the current state-of-the-art methods.

关 键 词：YOLOv8 遥感图像 小目标检测 特征融合 注意力机制 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]