数据与计算发展前沿 ›› 2024, Vol. 6 ›› Issue (6): 85-96.

CSTR: 32002.14.jfdc.CN10-1649/TP.2024.06.009

doi: 10.11871/jfdc.issn.2096-742X.2024.06.009

• • 上一篇    下一篇

基于联邦学习的野生动物红外相机图像目标检测

何文通1,2(),罗泽1,*()   

  1. 1.中国科学院计算机网络信息中心，北京 100083
    2.中国科学院大学，北京 100049
  • 收稿日期:2023-11-27 出版日期:2024-12-20 发布日期:2024-12-20
  • 通讯作者: 罗泽
  • 作者简介:何文通，中国科学院计算机网络信息中心，工程师，主要从事深度学习技术在生物多样性领域的应用研究。
    在本文中主要负责算法设计与实现，实验验证与文章撰写。
    HE Wentong is an engineer at the Computer Network Information Center, Chinese Academy of Sciences. His main research interests include the application of deep learning technology in the field of biodiversity.
    In this paper, he is mainly responsible for algorithm design and implementation, experimental vertification, and paper writing.
    E-mail: hwt0316@cnic.cn|罗泽，中国科学院计算机网络信息中心，研究员，主要研究方向为机器学习、科研信息化技术与应用。
    在本文中主要负责文章总体方向和框架。
    LUO Ze is a research fellow at the Computer Network Information Center, Chinese Academy of Sciences. His main research interests include machine learning and e-Science technologies and applications.
    In this paper, he is mainly responsible for the overall direction and framework of the paper.
    E-mail: luoze@cnic.cn
  • 基金资助:
    中国科学院网络安全与信息化专项(CAS-WX2022GC-0106)

Object Detection with Federated Learning for Wildlife Camera Trap Images

HE Wentong1,2(),LUO Ze1,*()   

  1. 1. Computer Network Information Center, Chinese Academy of Sciences, Beijing 100083, China
    2. University of Chinese Academy of Sciences, University, Beijing 100049, China
  • Received:2023-11-27 Online:2024-12-20 Published:2024-12-20
  • Contact: LUO Ze

RichHTML

7

PDF

231

摘要：

【目的】在数据保护日趋严格环境下，数据隐私和安全保护阻碍了深度学习技术在野生动物红外相机图像上的应用。【方法】本文首次利用开源联邦学习框架Flower和目标检测算法YOLOv8，在公开数据集上模拟真实场景下野生动物红外相机图像联邦学习，针对联邦学习参与方不同的本地学习轮数参数进行了对比实验，并和传统方式的深度学习进行对比。【结果】实验结果中不同本地训练轮数下联邦学习获得的全局目标检测模型mAP50最高可达到传统方式学习的95%，为保证数据隐私和安全，仅带来非常小的模型性能下降，表明联邦学习在野生动物红外相机图像深度学习领域具有非常大的应用潜力。又同各参与方的独立学习训练结果相比，联邦学习各参与方在学习训练过程中处于不公平地位，还需进一步研究适用于红外相机图像数据的联邦学习激励机制。【结论】本文在野生动物红外相机图像上的目标检测联邦学习实验表明联邦学习在野生动物红外相机图像上的应用研究对衡量和监测全球生物多样性变化具有重要意义。

关键词: 野生动物, 红外相机, 目标检测, 联邦学习, 深度学习

Abstract:

[Objective] In the increasingly strict environment of data protection, the data privacy and security has hindered the application of deep learning technology in wildlife camera trap images. [Methods] In this paper, we use the open-source federated learning framework Flower and object detection model YOLOv8 to simulate the federated learning of wildlife camera trap images in real scenes on the public dataset for the first time. Multiple experiments are conducted for different local learning epochs of federated learning clients, and compared with traditional deep learning methods. [Results] In the experimental results, the mAP50 of the global object detection models obtained by federated learning under different local epochs can reach 95% of that of traditional learning. To ensure data privacy and security, only a very small degradation in model performance is brought, indicating that federated learning has a very large application potential in the field of deep learning of wildlife camera trap images. Moreover, compared with the results of independent learning of each client of federated learning, each client is in an unfair position in federated learning, and the incentive mechanism for federated learning suitable for wildlife camera trap images needs to be further studied. [Conclusions] The multiple experiments of object detection with federated learning for wildlife camera trap images indicates that the application of federated learning on camera trap images has great significance for measuring and monitoring global biodiversity changes.

Key words: wildlife, camera traps, object detection, federated learning, deep learning