基于轻量化ViT的无锚框遥感影像目标检测方法  

Object Detection Method for Anchor-free Frame Remote Sensing Images Based on Lightweight ViT

作　　者：骆杰轩 LUO Jiexuan(Guangdong Provincial Institute of Land and Resources Surveying and Mapping,Guangzhou 510700,China)

机构地区：[1]广东省国土资源测绘院,广东广州510700

出　　处：《测绘与空间地理信息》2025年第3期117-120,123,共5页Geomatics & Spatial Information Technology

摘　　要：针对边缘计算场景下轻量模型对遥感影像目标检测精度较低的问题,提出一种融合高效视觉转换器与无锚点框机制的轻量级检测方法。在YOLOv5s基础上,在特征提取阶段引入高效视觉转换器网络,以轻量化结构同步捕获局部特征和上下文注意力权重;在特征融合阶段,以浅层特征图参与多尺度特征融合,并采用幽灵卷积核减少计算量;在检测阶段,引入无锚点框机制实施目标框位置回归,提高计算效率,并适应尺寸复杂的遥感影像目标;最后使用C++语言与TensorRT推理引擎,实现模型部署。结果表明,所提模型较YOLOv5s在检测精度、速度方面均有明显提升,能够在小算力环境下实时输出检测结果。In order to solve the problem of low detection accuracy of remote sensing image targets by lightweight models in edge computing scenarios,a lightweight detection method that combines an efficient visual converter and an anchor-free frame mechanism is proposed.Based on YOLOv5s,an Efficient Vision TransFormer network is introduced in the feature extraction stage to simultaneously capture local features and contextual attention weights with a lightweight structure;in the feature fusion stage,shallow feature maps are used to participate in multi-scale feature fusion,and ghost convolutional kernel is used to reduce the calculation amount;in the detection phase,an anchor-free frame mechanism is introduced to implement target frame position regression to improve calculation efficiency and adapt to remote sensing image targets with complex sizes;finally,the C++language and TensorRT inference engine are used to implement model deployment.The results prove that the proposed model has significantly improved detection accuracy and speed compared with YOLOv5s,and can output detection results in real time in a small computing power environment.

关 键 词：遥感影像 目标检测 高效视觉转换器 幽灵卷积核 无锚点框机制 

分 类 号：P237[天文地球—摄影测量与遥感]