现代防御技术 ›› 2022, Vol. 50 ›› Issue (1): 25-32.DOI: 10.3969/j.issn.1009-086x.2022.01.004

• 军事智能 • 上一篇    下一篇

一种RetinaNet与SE融合的航空取证目标检测算法

刘克1, 潘广煜2, 郑大国3, 顾佼佼1, 孟春英1   

  1. 1.海军航空大学, 山东 烟台 264001
    2.中国人民解放军92318部队, 北京 100000
    3.中国航天科工集团有限公司 第二研究院, 北京 100854
  • 收稿日期:2021-09-08 修回日期:2021-10-09 出版日期:2022-02-28 发布日期:2022-02-23
  • 作者简介:刘克(1984-),男,广西象州人。讲师,硕士,主要从事人工智能与大数据研究。
  • 基金资助:
    国家自然科学基金(62076249)

An Aerial Forensic Target Detection Algorithm Based on RetinaNet and SE Fusion

Ke LIU1, Guang-yu PAN2, Da-guo ZHENG3, Jiao-jiao GU1, Chun-ying MENG1   

  1. 1.Naval Aviation University, Coastal Defense College, Shandong Yantai 264001, China
    2.PLA, No. 92318 Troop, Beijing 100000, China
    3.The Second Academy of CASIC, Beijing 100854, China
  • Received:2021-09-08 Revised:2021-10-09 Online:2022-02-28 Published:2022-02-23

摘要:

针对在开展航空侦察取证时缺乏自动化、智能化取证手段的情况,提出了一种基于RetinaNet与SE融合的航空取证目标检测算法,解决目标尺度变化大、数据集中类别不均衡的问题,并通过SE(squeeze-excitations)模块引入注意力机制进一步改进性能。该改进算法中的特征金字塔网络FPN可有效应对目标尺寸变化较大问题,Focal Loss可有效应对数据均衡性问题,SE模块引入通道注意力机制对特征图进行加强,可进一步利用提取的通道间相关性增强有效特征并抑制无效特征。通过仿真实验,验证了算法能够在增加少量计算条件下提高目标检测准确率,进一步增强模型的表征能力,有效提高目标检测效率,可为相关工程应用提供参考。

关键词: 航空取证, 目标检测, RetinaNet, SE模块, 数据均衡, 通道注意力

Abstract:

Aiming at the lack of automated and intelligent forensics methods in the aviation reconnaissance and evidence collection, an aviation forensics target detection algorithm based on the fusion of RetinaNet and SE is proposed to solve the problem of large changes in target scale and imbalance in data collection. The performance is further improved by introducing an attention mechanism through the SE (squeeze-excitations) module. The feature pyramid network(FPN) can effectively deal with the problem of large changes in target scale, Focal Loss can effectively deal with the imbalance of categories in dataset, the SE module introduces a channel attention mechanism to strengthen the feature map, and can further use the extracted channel correlation to enhance effective features and suppress ineffective features. Through simulation experiments, it is verified that the algorithm can improve the accuracy of target detection with a small amount of calculation increased, further enhance the characterization ability of the model, and effectively improve the efficiency of target detection, which can provide a reference for related engineering applications.

Key words: aviation forensics, target detection, RetinaNet, Squeeze-excitation(SE) module, data equalization, channel attention

中图分类号: