半监督式野生动物夜间目标端到端检测  

End-to-end recognition of nighttime wildlife based on semi-supervised learning

作　　者：卢汉 崔博伦 万华洋 张国峰 沈晨 王驰 LU Han;CUI Bolun;WAN Huayang;ZHANG Guofeng;SHEN Chen;WANG Chi(School of Mechatronic Engineering and Automation,Shanghai University,Shanghai 200444,China;Beijing Institute of Space Mechanics&Electricity,Beijing 100094,China)

机构地区：[1]上海大学机电工程与自动化学院,上海200444 [2]北京空间机电研究所,北京100094

出　　处：《光学精密工程》2025年第5期789-801,共13页Optics and Precision Engineering

基　　金：北京市航空智能遥感装备工程技术研究中心开放基金课题(No.AIRSE20233);国家重点研发计划资助项目(No.2023YFF0722902);国家自然科学基金资助项目(No.62175144)。

摘　　要：针对野生动物夜间目标检测精度与效率低且难以人工全面标注的问题,提出了一种基于半监督学习的夜间野生动物端到端识别模型(SAN-YOLO)。基于YOLOv8融合特征注意力机制和像素注意力机制,以提升检测器对于夜间图像的自适应性和特征表征能力。然后,搭建了基于师生学习框架的半监督训练网络,通过生成并合理分配伪标签,让学生模型从大量未标注的原始图像中学习,最后,在所构建的数据集上进行评估。实验结果表明,在使用5%标注的情况下,SAN-YOLO的mAP指标达到69.7%,高于其检测器在全监督训练下的59.6%和基线模型的57.1%。所研究的检测方法在少量标注的夜间动物数据集上展现出良好的性能,且验证了注意力机制在夜间目标检测领域的有效性。This study addresses the challenges of low accuracy and efficiency in the detection of wildlife at night,as well as the difficulties associated with manual comprehensive labeling.An end-to-end recognition model for nighttime wildlife based on semi-supervised learning(SAN-YOLO)was proposed and investigated.A feature attention mechanism and a pixel attention mechanism were integrated within the YO-LOv8 framework to enhance the adaptability and feature representation capabilities of the detector for nocturnal images.Subsequently,a semi-supervised training network based on a teacher-student learning paradigm was constructed,allowing the student model to learn from a substantial number of unlabeled original images by generating and appropriately assigning pseudo-labels.The efficacy of the constructed dataset was then evaluated.Experimental results demonstrate that the mean Average Precision(mAP)of SAN-YOLO reaches 69.7%with only 5%annotated data,surpassing the 59.6%mAP achieved with full su⁃pervision in its conventional detector and exceeding the baseline model's performance of 57.1%.Consequently,the proposed detection method exhibits robust performance with a limited number of labeled datasets for nocturnal animals and validates the effectiveness of attention mechanisms in the domain of nighttime object detection.

关 键 词：目标检测 半监督学习 红外夜视 野生动物保护 师生模型 注意力机制 

分 类 号：P751.1[交通运输工程—港口、海岸及近海工程]