基于改进EAF-YOLO的轻量化遥感图像检测  

Lightweight Remote Sensing Image Detection Based on Improved EAF-YOLO

作　　者：侯林静 李芳[1] HOU Linjing;LI Fang(Shenyang Ligong University,Shenyang 110159,China)

机构地区：[1]沈阳理工大学信息科学与工程学院,沈阳110159

出　　处：《沈阳理工大学学报》2025年第2期7-12,共6页Journal of Shenyang Ligong University

基　　金：国家自然科学基金项目(62102272)。

摘　　要：针对遥感图像目标检测模型难以在无人机、卫星以及移动设备等低算力场景下部署的问题,提出一种轻量化遥感图像检测方法EAF-YOLO。模型以YOLOv7框架为基础,采用Ef-ficientNet特征提取网络作为轻量级骨干网络,降低模型参数;在框架头部设计串行结构的空间金字塔池化层,提升内部尺度交换能力,增强对特征信息的聚合能力;采用WIoU损失函数作为回归损失,自适应地调整难易样本的权重,提高模型的检测性能。以遥感数据集DIOR进行实验,EAF-YOLO模型检测精确率达到了90.3%,参数量为9.3×10^(6),表明训练后的模型较YOLOv7具有较小的参数量,适用于遥感图像计算资源有限的场景。To address issues of deploying remote sensing image object detection models in low computing scenarios such as drones,satellites,and mobile devices,a novel method utilizing EAFYOLO for lightweight remote sensing image detection is proposed.The model is based on the YOLOv7 framework.The EfficientNet feature extraction network is utilized as a lightweight backbone network to diminish the model parameters.A spatial pyramid pooling layer is crafted with a consecutive configuration within the Head section,improving the scale interactions capability and fortifying the aggregation capacity of feature information.The model employs the WIoU loss function as the regression loss,allowing for the adaptive adjustment of weights for difficult samples to enhance the detection performance.Experimenting with the remote sensing dataset DIOR reveals that,the EAF-YOLO model achieves an precision of 90.3%,while its parameter count decreases to 9.3×10^(6).Due to its low parameter count and computational intricacy,it is apt for utilization in remote sensing scenarios where there are constraints on computing resources.

关 键 词：遥感图像检测 轻量化 YOLOv7 EfficientNet 尺度交换 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]