摘要
【目的】设计并实现一种基于YOLOv8的鱼苗计数算法,提升计数的自动化水平和检测精度,以解决传统人工鱼苗计数效率低、误差大、难以满足规模化养殖管理需求等问题。【方法】基于YOLOv8目标检测框架,增加形态特征提取、生存状态识别和重叠分割优化3个后处理模块,实现对模糊目标、低活性个体和重叠鱼苗的检测结果校正与精度增强。【结果】相较YOLOv8,改进算法对2~4 cm活体草鱼(Ctenopharyngodon idellus)苗的计数准确率由84.6%提升至92.1%,平均绝对误差(MAE)由0.42尾/帧降至0.15尾/帧,平均绝对百分比误差(MAPE)由3.2%降至1.4%,有效降低因目标模糊、失去活性或重叠造成的误差。稳定性测试表明,计数准确率变异系数为1.6%。在斑马鱼(Danio rerio)、杂交剑尾鱼(Xiphophorus hellerii×Xiphophorus maculatus)和科斯塔氏直线脂鲤(Moenkhausia costae)3类鱼苗实验中,平均计数准确率由83.8%提升至92.0%,平均MAE由0.44尾/帧降至0.15尾/帧,平均MAPE由3.03%降至1.36%,改进算法的稳定性与泛化能力得到验证。【结论】基于YOLOv8的改进算法能够实现鱼苗的连续、高效、精准计数,具有较好的泛化能力。
【Objective】This study aims to design and implement an automated fish fry counting system based on YOLOv8,improve the automation level and detection accuracy of counting,addressing the problems of low efficiency and high error rates in traditional manual fry counting methods,which cannot meet the needs of large-scale aquaculture management.【Method】Based on the YOLOv8 object detection framework,three post-processing modules including morphological feature extraction,survival state recognition and overlapping segmentation optimization were added to achieve correction and accuracy enhancement of detection results for fuzzy targets,low activity individuals,and overlapping fish fry.【Result】Compared to YOLOv8,the algorithm improved the counting accuracy of 2 to 4 cm live fish fry of Ctenopharyngodon idellus from 84.6%to 92.1%,reduced mean absolute error(MAE)from 0.42 fish/frame to 0.15 fish/frame,and mean absolute percentage error(MAPE)from 3.2%to 1.4%,effectively reducing errors caused by target blurring,loss of activity,or overlap.Stability testing showed that the coefficient of variation for counting accuracy was 1.6%.In the experiments of three types of fish fry,Danio rerio,Xiphophorus hellerii×Xiphophorus maculatus,and Moenkhausia costae,the average counting accuracy increased from 83.8%to 92.0%,the average MAE decreased from 0.44 fish/frame to 0.15 fish/frame,and the average MAPE decreased from 3.03%to 1.36%.The stability and generalization ability of the improved algorithm were verified.【Conclusion】Improved algorithms based on YOLOv8 can achieve continuous,efficient,and accurate counting of fish fry,with good generalization ability.
作者
杨棋贺
姜忠爱
于赢水
王鹏
吴俊峰
李双双
YANG Qihe;JIANG Zhongai;YU Yingshui;WANG Peng;WU Junfeng;LI Shuangshuang(College of Mechanical and Power Engineering,Dalian Ocean University,Dalian 116023,China;Dalian Key Laboratory of Intelligent Fisheries,Dalian 116023,China;Dalian Xinyulong Marine Biological Seed Industry Technology Co.,Ltd.,Dalian 116222,China)
出处
《广东海洋大学学报》
北大核心
2025年第6期132-139,共8页
Journal of Guangdong Ocean University
基金
国家重点研发计划“政府间国际科技创新合作”项目(2022YFE0117900)
设施渔业教育部重点实验室(大连海洋大学)开放课题(202315)
辽宁省教育厅科研项目(JYTMS20230501,JYTMS20230502)。
关键词
鱼苗计数
YOLOv8
计数效率
准确率
形态特征提取
生存状态识别
重叠分割优化
fish fry counting
YOLOv8
counting efficiency
accuracy
morphological feature extraction
survival status recognition
overlap segmentation optimization
