基于对抗学习的野生动物图像域适应识别方法  

A Recognition Method of Domain Adaptation for Wildlife Images Based on Adversarial Learning

作　　者：赵恩庭 张长春 赵海涛[2] 张军国[1] Zhao Enting;Zhang Changchun;Zhao Haitao;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;Shaanxi Institute of Zoology,Xi’an 710032)

机构地区：[1]北京林业大学工学院林木资源高效生产全国重点试验室、林业装备与自动化国家林业和草原局重点实验室,北京100083 [2]陕西省动物研究所,西安710032

出　　处：《林业科学》2025年第4期1-8,共8页Scientia Silvae Sinicae

基　　金：北京市自然科学基金项目(6244053);国家自然科学基金项目(32371874,32401569)。

摘　　要：【目的】针对不同时空场景、物种差异等引发的域偏移问题,提出一种基于对抗学习的野生动物图像域适应识别方法,提升复杂野外环境下无标签野生动物物种识别的泛化性能,为开放环境野生动物分类研究提供有效理论依据。【方法】利用野生动物图像的类别信息作为条件构建对抗学习网络模型,通过批光谱惩罚约束和mixup特征对齐方法减轻不同时空场景、物种差异下野生动物图像间的分布差异,建立基于对抗学习的野生动物图像域适应模型实现图像识别。【结果】在分别包含8和11种野生动物的域适应数据集上训练和评估本研究提出的方法,与联合对抗学习和特征对齐的野生动物识别基线方法相比,本研究提出方法的平均识别准确率分别提升3.3%、14.0%,精确率分别提升3.3%、20.6%,召回率分别提升3.5%、20.5%,F1值分别提升3.6%、5.1%,基于对抗学习的野生动物图像域适应模型对不同时空场景下野生动物的识别性能明显提升。【结论】野生动物图像的类别信息作为对抗学习网络条件,可提供野生动物图像的多模态结构信息,有助于本研究方法更好理解野生动物图像之间的关系,提升野生动物图像域适应学习性能。训练集和测试集图像特征对齐得越好,测试集的图像识别性能越好。本研究为野生动物图像跨域识别研究提供了新的思路和方法。【Objective】To address the domain shift issues caused by different spatiotemporal scenarios and species variations,this study proposes a domain adaptation recognition method for wildlife images based on adversarial learning.The aim of this study is to enhance the generalization performance of unlabeled wildlife species recognition in complex field environment,thereby providing a robust theoretical foundation for the classification of wildlife in open environment.【Method】An adversarial learning network model was constructed by utilizing wildlife image category information as conditional inputs.Batch spectral penalty constraints and mixup feature alignment methods were employed to mitigate distribution discrepancies among wildlife images across different spatiotemporal scenarios and species variations.This framework established a feature alignment model based on adversarial learning to enhance image recognition performance.【Result】The proposed method was trained and evaluated on domain adaptation datasets containing 8 and 11 wildlife species,respectively.Compared to baseline methods combining adversarial learning and feature alignment,the average recognition accuracy of the proposed method increased by 3.3%and 14.0%,the precision was improved by 3.3%and 20.6%,the recall rate was improved by 3.5%and 20.5%,and the F1-score was improved by 3.6%and 5.1%,respectively.The wild animal image domain adaptation model based on adversarial learning significantly improved performance for wildlife recognition in cross-domain scenarios.【Conclusion】The wildlife category information used as conditional inputs in the adversarial learning network can provide multimodal structural information of wildlife images,which enables the proposed method in this study to better understand the inter-image relationships and improve domain adaptation learning capabilities.The better the alignment of image features between the training and testing sets,the better the image recognition performance of the testing set.This study provide

关 键 词：野生动物 识别方法 对抗学习 特征对齐 域适应 

分 类 号：S718.6[农业科学—林学] S862