摘要
【目的】针对林火烟雾图像识别中因时空场景差异引起的域偏移问题,本研究提出一种基于域对抗网络的林火烟雾图像跨域识别方法。该方法旨在解决现有技术在跨域场景下识别性能不足的问题,特别是在复杂背景和不同环境条件下的应用挑战,为复杂环境下的森林火灾监测提供技术支撑。【方法】首先,本研究方法在训练阶段引入了条件对抗学习机制,利用林火烟雾图像的类别信息构建条件约束网络,增强模型对跨域特征的适应能力。然后,设计域不变特征提取模块和跨域特征对齐模块。在域不变特征提取模块中,采用预训练的ResNet50作为基础架构,生成具有判别性的域不变特征,解决跨域场景下的特征分布差异问题。在跨域特征对齐模块中,融合最大均值差异度量和相关对齐约束,构建双重约束机制以优化特征空间分布。最终,通过融合条件对抗生成网络与跨域特征对齐模块,构建端到端的训练框架,实现林火烟雾图像的跨域高效识别。【结果】在相同实验条件下,本文所提方法在跨域林火烟雾数据集中平均识别准确率达92.39%,较最优基线模型(LEAD)提升了0.94个百分点,且精确率、召回率和F1值这3个关键指标分别达到了89.67%、89.58%和89.54%,均显著优于现有的方法。这一结果充分验证了本方法对提升不同时空场景下林火烟雾识别性能方面的有效性。多场景泛化实验结果表明,在复杂气象的干扰下,该模型的识别性能仍较为稳定。【结论】综上所述,本文提出的基于域对抗网络的林火烟雾图像跨域识别方法,有效提升了不同时空场景下林火烟雾图像的识别性能,展现了在复杂环境下的高鲁棒性。该方法为林火烟雾图像跨域识别研究提供了新的思路和解决方案,为相关领域的进一步探索提供了通用参考价值。
[Objective]This study suggests a cross-domain identification technique for forest fire smoke photos based on domain adversarial networks in order to address the issue of domain shift caused by spatiotemporal scene changes in forest fire smoke image detection.The approach attempts to address the issue of current methods’inadequate recognition performance in cross-domain scenarios,particularly the difficulties in using them in complicated environments with varying backgrounds.It offers technical assistance for monitoring forest fires in intricate settings.[Method]Firstly,this study introduced a conditional adversarial learning mechanism in training stage,which utilized the category information of forest fire smoke images to construct a conditional constraint network and enhanced the model’s ability to adapt to cross-domain features.Then,the domain invariant feature extraction module and the cross-domain feature alignment module were designed.In the domain-invariant feature extraction module,the pre-trained ResNet50 was used as infrastructure to generate discriminative domain-invariant features to solve the problem of feature distribution differences in cross-domain scenarios.In the cross-domain feature alignment module,the maximum mean difference metric and associated alignment constraints were fused to construct a dual constraint mechanism to optimize the feature spatial distribution.Finally,by fusing the conditional adversarial generative network and cross-domain feature alignment module,an end-to-end training framework was constructed to realize the cross-domain efficient recognition of forest fire smoke images.[Result]The suggested approach’s average recognition accuracy in cross-domain forest fire smoke dataset under same experimental conditions was 92.39%,0.94 percentage points higher than optimal baseline model(LEAD).Additionally,the key metrics(89.67%precision,89.58%recall,and 89.54%F1 score)were significantly better than the others,confirming that the approach presented in this paper was successful in enhancing the performance of forest fire smoke recognition.Experiments on multi-scene generalization demonstrated that even in the face of complicated meteorological interference,the model’s identification ability remained comparatively constant.[Conclusion]In conclusion,the domain adversarial networkbased cross-domain recognition technique for forest fire smoke images presented in this paper successfully enhanced the recognition performance of forest fire smoke images in various spatiotemporal scenarios and exhibited high robustness in complex environments.The approach offers fresh concepts and methods for studying forest fire smoke picture cross-domain identification,as well as a general reference value for additional research in related disciplines.
作者
赵雨诺
张长春
张军国
Zhao Yunuo;Zhang Changchun;Zhang Junguo(School of Technology,Beijing Forestry University,State Key Laboratory of Efficient Production of Forest Resources,Key Laboratory of National Forestry and Grassland Administration on Forestry Equipment and Automation,Beijing 100083,China)
出处
《北京林业大学学报》
北大核心
2025年第6期130-140,共11页
Journal of Beijing Forestry University
基金
国家重点研发计划项目(SQ2023XAGG0065)
中央高校优秀青年团队项目(QNTD202304)。
关键词
林火烟雾
图像识别
跨域识别
对抗学习
特征对齐
迁移学习
smoke of forest fire
image recognition
cross-domain recognition
adversarial learning
feature alignment
transfer learning
作者简介
第一作者:赵雨诺。主要研究方向:智慧林业与森林防火。Email:18342819268@163.com,地址:100083北京市海淀区清华东路35号北京林业大学工学院;张长春,讲师。主要研究方向:人工智能与模式识别、迁移学习。Email:zhangchangchun@bjfu.edu.cn,地址:100083北京市海淀区清华东路35号北京林业大学工学院;张军国,教授。主要研究方向:图像处理以及深度学习、林草生态智慧监测。Email:zhangjunguo@bjfu.edu.cn,地址:100083北京市海淀区清华东路35号北京林业大学工学院。
