空面导弹轻量化空中斜框目标检测算法  

A Lightweight Aerial Rotated Object Detection Algorithm of Air-to-Surface Missile

作　　者：刘婧 郭晓雷 张欣海 毛靖军 吕瑞恒 LIU Jing;GUO Xiaolei;ZHANG Xinhai;MAO Jingjun;LYU Ruiheng(School of Electrical and Information Engineering,Tianjin University,Tianjin 300072,China;China Academy of Electronics and Information Technology,Beijing 100041,China;Shanghai Electro-Mechanical Engineering Institute,Shanghai 201109,China)

机构地区：[1]天津大学电气自动化与信息工程学院,天津300072 [2]中国电子科技集团有限公司电子科学研究院,北京100041 [3]上海机电工程研究所,上海201109

出　　处：《空天防御》2024年第4期106-113,共8页Air & Space Defense

基　　金：国家自然科学基金面上项目(62371333)。

摘　　要：随着现代军事和安全挑战的演变,地面目标种类和方向不断增加,传统的目标检测算法在复杂多变的空中环境时存在诸多限制,且因神经网络的复杂性使其在移动设备上难以应用。因此,为了提升飞行过程中对地面任意方向目标检测的准确率和检测效率,提出轻量化空中斜框目标检测算法。首先,使用Ghost模块改进原有的卷积神经网络,利用少量小滤波器生成更多的特征图,减少计算成本;其次,使用改进后的卷积网络从特征金字塔的5个级别的特征中生成高质量候选目标框;最后,使用检测头接收生成的候选目标框作为输入,计算得分后最终用于分类和回归。实验结果表明,本文所提出的斜框目标检测算法能够对地面目标进行精确的检测定位并降低了模型复杂度。With the evolution of modern military and security challenges,the variety and direction of ground targets increase.Traditional target detection algorithms are facing numerous limitations in complex and dynamic aerial environments,and the complexity of neural networks challenges their application on mobile devices.Therefore,to enhance the accuracy and efficiency of detecting ground targets from any direction during flight,a lightweight aerial rotated object detection algorithm(LRODA)was proposed in this study.Firstly,the Ghost module was employed to enhance the original convolutional neural network and generated more feature maps with fewer small filters,thus reducing computational costs.Secondly,the improved convolutional network was used to generate high-quality proposals from the feature pyramid network at five different levels.Finally,the detection head harvested the generated proposals as input and ultimately used for classification and regression.Experimental results prove that the proposed rotated object detection algorithm can locate ground targets accurately and reduce model complexity effectively.

关 键 词：斜框目标检测 轻量化学习 特征金字塔 深度学习 

分 类 号：TP391[自动化与计算机技术—计算机应用技术] TP18[自动化与计算机技术—计算机科学与技术]