基于YOLOv5目标检测算法的航道工程鱼类智能识别技术研究  

Study on intelligent identification based on object detection algorithm of YOLOv5for fish in waterway engineering

作　　者：邱宁 彭士涛[1] KONDOLF George Mathias 胡健波[1] 马国强 于迅 QIU Ning;PENG Shitao;KONDOLF George Mathias;HU Jianbo;MA Guoqiang;YU Xun(Tianjin Research Institute for Water Transport Engineering,National Engineering Research Center of Port Hydraulic Construction Technology,Key Laboratory of Environmental Protection Technology on Water Transport,Ministry of Transport,Tianjin 300456,China;University of California,Berkeley,Berkeley 48072,USA)

机构地区：[1]交通运输部天津水运工程科学研究所港口水工建筑技术国家工程研究中心水路交通环境保护技术交通行业重点实验室,天津300456 [2]加州大学伯克利分校,伯克利48072

出　　处：《水道港口》2024年第5期806-815,共10页Journal of Waterway and Harbor

基　　金：国家重点研发计划政府间国际科技创新合作项目(2019YFE0121000);国家自然科学基金项目(32101241);青海省交通运输厅科技项目(2019-07)。

摘　　要：为促进深度学习技术在航道工程生态影响评估方面的应用,提高内河航道建设的生态化和智能化水平。文章筛选了36种长江航道工程区域典型鱼类为目标鱼种,采用原位与室内两种场景水下影像采集及人工标注的方法,建立了航道工程鱼类目标检测数据集。进一步利用该数据集基于YOLOv5目标检测算法进行了模型训练,并对训练的模型进行了测试验证。模型测试结果表明,训练数据集的精确率分数为0.933,召回率分数为0.98,平衡点处F1分数为0.89,均接近于1,训练效果较好。训练数据集与验证数据集的标注框损失值、目标损失值和分类损失值均趋近于0,训练和验证数据的拟合程度良好。混淆矩阵图表明不同鱼体之间能够互相区分,且能够预测为各自正确的类别。验证数据集整体mAP值为0.933,召回率分数0.98,F1平衡点处分数0.89,识别效果较好。整体测试结果表明研究形成的基于YOLOv5的目标检测技术对航道工程典型鱼类具有优良的目标检测与识别效果。To promote the application of deep learning technology in the ecological impact assessment of waterway engineering and enhance the ecological and intelligent levels of inland waterway construction,36 typical fish species in the Yangtze River waterway engineering area were selected in this paper as target species.A fish target detection dataset for waterway engineering was established by adopting underwater image acquisition and manual annotation methods in both in-situ and indoor scenarios.Furthermore,a model based on the YOLOv5 object detection algorithm was trained by above dataset,which was subsequently tested and validated.The model test results indicated that the precision score of the training dataset was 0.933,the recall score was 0.98,and the F1 score at the balance point was 0.89,all close to 1,demonstrating effective training outcomes.The bounding box loss,object loss,and classification loss values of both the training and validation datasets approached zero,indicating a good fit between training and validation data.The confusion matrix diagram showed that different fish bodies could be distinguished from each other and accurately predicted into their respective categories.The overall mAP value of the validation dataset was 0.933,with a recall score of 0.98 and an F1 score of 0.89 at the balance point,indicating excellent recognition performance.Overall,the test results demonstrate that the YOLOv5-based object detection technology developed in this study achieves outstanding target detection and recognition effects for typical fish species in waterway engineering.

关 键 词：YOLOv5 目标检测 鱼类智能识别 航道工程 深度学习 生态影响评估 

分 类 号：U65[交通运输工程—港口、海岸及近海工程] P229[交通运输工程—船舶与海洋工程]