复杂场景下军事目标的轻量级检测方法  

Lightweight detection method for military targets in complex scenes

作　　者：刘茹茹[1] 洪锋[1,2] 孙佐 LIU Ru-rul;HONG Feng;SUN Zuo(School of Mechanical and Electrical Engineering,Chizhou University,Chizhou Anhui,247000,China;School of Computer and Information,Hefei University of Technology,Hefei,Anhui 236000,China)

机构地区：[1]池州学院机电工程学院,安徽池州247000 [2]合肥工业大学计算机与信息学院,安徽合肥236000

出　　处：《光电子．激光》2021年第5期541-548,共8页Journal of Optoelectronics·Laser

基　　金：安徽省优秀青年人才培育计划项目(gxyq2018110,No.gxyq2019111);池州学院自然重点项目(cz2019zrz07)资助项目。

摘　　要：人工智能技术在现代战争中具有举足轻重的地位,复杂环境下的军事目标精准识别有利于使我方抢占先机从而克敌制胜。秉承海空一体化联合作战的重要理念,为使目标检测方法能够嵌入到各种军事单位中的微电脑中,提出了一种轻量且精准的军事目标检测方法。通过分析与结合士兵、汽车、与坦克及其履带的特征,设计出一个轻量级的网络单元。使用网络单元组成一种计算复杂度低且精准的骨干网络,用于提取目标特征信息。设计一个MSCA(multi-scale context aggregation)模块,对骨干网络中高维与低维的特征分别提取,解决了目标遮挡的问题。实验结果表明,本文方法在军事目标测试集中的准确识别率为97.8%,与最新的YOLOv4检测方法相比,检测准确度提高了1.1%,运行速度提高了5倍,能够满足嵌入式设备实时运行的要求。通过实验可得,本文方法可以实时且精准的检测多种场景下的军事目标。Artificial intelligence technology has a pivotal position in modern warfare.The precise identification of military targets in a complex environment will help us seize the opportunity to defeat the enemy.Adhering to the important concept of sea air integrated joint operation,a lightweight and accurate military target detection method is proposed in order to embed the target detection method into the microcomputers of various military units.By analyzing and combining the characteristics of soldiers,cars,tanks and their tracks,a lightweight network unit is designed.The network unit is used to form a backbone network with low computational complexity and accuracy,which is used to extract the target feature information,which improves the detection speed of the method.A MSCA(multi-scale context aggregation)module is designed to improve the high-dimensional performance of the backbone network.Separate extraction from low-dimensional features solves the problem of target occlusion.Experimental results show that the accurate recognition rate of this method in the military target test set is 97.8%.Compared with the YOLOv4 detection method,the detection accuracy is improved by 1.1%,and the running speed is increased by 5 times,which can meet the requirements of real-time operation of embedded devices.Through experiments,the proposed method can detect military targets in a variety of scenarios in real time and accurately.

关 键 词：图像识别 军事目标检测 嵌入式平台 轻量级网络 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]