ReNet:基于无锚框的地面旋转目标检测方法  

ReNet:Ground Rotating Target Detection Method Based on Anchor-Free Frame

作　　者：徐叶斌 王云鹏 刘少龙 刘力 李瑞 XU Yebin;WANG Yunpeng;LIU Shaolong;LIU Li;LI Rui(Xi'an Aviation Computing Technique Research Institute,AVIC,Xi'an 710065,China)

机构地区：[1]中国航空工业集团公司西安航空计算技术研究所,陕西西安710065

出　　处：《红外技术》2025年第2期211-216,共6页Infrared Technology

基　　金：航空科学基金(2022Z071031001)。

摘　　要：地面红外目标检测是高空侦察、智能感知和对地打击等领域的重要研究内容,针对所获取的地面红外目标常以不规则角度的形式出现,导致检测准确率低,容易发生误检、漏检等问题。以Anchor-Free目标检测模型为基础,构建了基于空洞卷积为特征提取方式的主干网络,增强了模型对地面旋转目标的感知范围与特征提取能力;在基于空洞卷积进行特征提取后,通过External attention(EA注意力机制)增加对所提取特征注意维度的关注,实现了对目标更高分辨率特征的提取,最终提出了基于无锚框(Anchor-Free)的地面旋转目标检测方法。构建的地面旋转目标检测模型在HIT-UAV数据集上达到了90.6%的检测精度,优化了基于Anchor-Free的目标检测模型针对地面旋转目标的检测性能。Ground infrared target detection is crucial in the fields of high-altitude reconnaissance,intelligent perception,and ground strike,where the acquired ground infrared targets often appear in the form of irreg-ular angles,resulting in low detection accuracy,ease of misdetection,and other problems.Therefore,this paper proposes an anchor-free-based ground rotating target detection method.Based on the anchor-free tar-get detection model,a backbone network based on atrous convolution is constructed,which enhances the perception range and feature extraction ability of the model for ground rotating targets.After feature extrac-tion based on void convolution,the attention dimension of the extracted feature is increased through external attention,and the extraction of higher-resolution features of the target is realized.The ground rotating target detection model achieved 90.6%detection accuracy on the HIT-UAV dataset,which optimized the detection performance of the anchor-free target detection model for ground rotating targets.

关 键 词：深度学习 地面旋转目标检测 智能感知 Anchor-Free 空洞卷积 EA注意力机制 

分 类 号：TP371[自动化与计算机技术—计算机系统结构]