基于改进YOLOv5的军事目标识别方法  

A Military Target Detection Method Based on Improved YOLOv5

作　　者：万晓刚 王伟[1] WAN Xiaogang;WANG Wei(School of Computer Science,Xi'an Polytechnic University,Xi'an 710600)

机构地区：[1]西安工程大学计算机科学学院,西安710600

出　　处：《舰船电子工程》2024年第4期28-33,共6页Ship Electronic Engineering

基　　金：2021年中国高校产学研创新基金项目(编号:2021ALA02002)资助。

摘　　要：针对战场环境下因背景干扰和军事目标尺度较小等原因导致误检、漏检的问题,提出一种基于改进YOLOv5的军事目标识别方法 CB-YOLOv5。利用坐标注意力机制重构特征提取主干网络,增强网络对复杂背景下军事目标的特征提取能力;在特征融合网络中引入BiFPN,减少浅层特征信息的丢失,提高对弱小目标的检测能力。在自建数据集下实验表明,改进后算法mAP达到93.8%,比原模型提升了3.5%,可以有效识别战场环境下多尺度军事目标。To address the problem of false detection and missed detection due to background interference and small scale of military targets in battlefield environment,a military target recognition method CB-YOLOv5 based on improved YOLOv5 is pro⁃posed.The feature extraction backbone network is reconstructed by using coordinate attention mechanism to enhance the feature ex⁃traction capability of the network for military targets in complex background.BiFPN is introduced in the feature fusion network to re⁃duce the loss of shallow feature information and improve the weak targets can be detected.Experiments under the self-built dataset show that the improved algorithm mAP reaches 93.8%,which is 3.5%better than the original model,and can effectively identify multi-scale military targets in the battlefield environment.

关 键 词：目标识别 YOLOv5 注意力机制 特征融合 

分 类 号：TP301.6[自动化与计算机技术—计算机系统结构]