摘要
传统目标检测算法在河道漂浮物检测中存在类别局限性,无法满足精细化检测需求。为此,提出了一种基于深度学习的多技术融合方案,首先采用DeepLabv3+语义分割技术分离出河道水岸,消除岸域背景干扰;然后结合Detic万物检测,通过自定义词汇表将检测类别从几十种扩展至上万种;最后引入黑白名单机制对检测到的候选漂浮物进行过滤、排除非漂浮目标,以聚焦于治理部门关注的漂浮物类别。实验结果表明:该方法在保证检测效率的同时,类别识别准确率达87.8%,较YOLOv10算法提升4.8%,比YOLOv12算法提升1.7%;该方法支持边缘设备部署,推理速度为0.179 s/帧,能够满足河道复杂场景下的精准治理需求。
The traditional target detection algorithms face limitations of categories when they identify river floating objects,which fails to meet current demands for fine-grained detection.To address this,a multi-technology fusion method based on deep learning is proposed.First,DeepLabv3+semantic segmentation is applied to separate the riverbank and eliminate background interference from the shore.Then,by integrating Detic's open-vocabulary detection capability,the range of detectable categories is expanded from dozens to tens of thousands through a custom vocabulary.Finally,a whitelist-blacklist mechanism is introduced to filter candidate floating objects,excluding non-floating targets and focusing on the types of floatables that are relevant to environmental management.Experimental results show that while ensuring detection efficiency,the proposed method achieves a category recognition accuracy of 87.8%,outperforming the YOLOv10 algorithm by 4.8% and the YOLOv12 algorithm by 1.7%,thereby realizing fine-grained detection of river floating objects.Furthermore,the method supports deployment on edge devices,with an inference speed of 0.179 seconds per frame,meeting the requirements for accurate management in complex river environments.
作者
刘芳芳
王红
李丹
李博霖
边彬彬
LIU Fangfang;WANG Hong;LI Dan;LI Bolin;BIAN Binbin(Beijing Institute of New Technology Applications,Beijing 100089,China)
出处
《人民长江》
北大核心
2025年第12期264-270,共7页
Yangtze River
基金
新疆维吾尔自治区重点研发计划项目(2022B01008)
北京市西城区优秀人才项目(202339)。
关键词
河道漂浮物
目标检测
河道水岸分割
Detic万物检测
自定义词汇表
黑白名单
river floating objects
target detection
riverbank segmentation
Detic everything detection
custom vocabulary
black-and-white list
