基于改进YOLOv7-tiny的凡纳滨对虾游动活跃性定量检测方法  

Quantitative detection method of swimming activity of Litopenaeus vannamei based on improved YOLOv7-tiny

作　　者：李志坚 张永琪 吴迪 孟雄栋[1,2] 李延天 张丽珍 LI Zhijian;ZHANG Yongqi;WU Di;MENG Xiongdong;LI Yantian;ZHANG Lizhen(College of Engineering Science and Technology,Shanghai Ocean University,Shanghai 201306,China;Shanghai Engineering Research Center of Marine Renewable Energy,Shanghai Ocean University,Shanghai 201306,China)

机构地区：[1]上海海洋大学工程学院,上海201306 [2]上海海洋大学,上海海洋可再生能源工程技术研究中心,上海201306

出　　处：《水产学报》2024年第12期83-94,共12页Journal of Fisheries of China

基　　金：国家重点研发计划(2019YFD0900401);上海市水产动物良种创制与绿色养殖协同创新中心项目(2021科技02-12)。

摘　　要：在凡纳滨对虾养殖过程中,养殖人员需通过人为提拉饲料盘来观测对虾的游动活跃状态,以此了解其生长状况并制定投喂策略。为了解决人工观察凡纳滨对虾游动活跃状态时存在实时性不佳、主观因素影响大等问题,提出一种基于改进YOLOv7-tiny (improved YOLOv7-tiny)检测模型和基于欧式距离多目标关联视觉检测方法,用以定量分析饲料盘上对虾游动活跃状态。在YOLOv7-tiny模型基础上,将标准卷积替换为GSConv卷积,搭建VoVGSCSPC模块替换原先轻量化聚合(ELAN-L)模块并使用损失函数MPDIoU替代原损失函数CIoU,减少了模型容量并提升模型检测精度。通过改进YOLOv7-tiny模型的检测结果和基于欧式距离的多目标关联方法确定图像中对虾的位置,据此计算对虾的游动位移、速度与转角,量化出对虾游动活跃状态。在凡纳滨对虾数据集上验证后,结果显示,与YOLOv7-tiny模型相比,改进YOLOv7-tiny模型错检率和漏检率分别降低0.62%与1.05%,推理速度提升17.07%,验证了改进后的模型有效性。通过对虾游动活跃性进行定量分析发现,越活跃的对虾活跃性指标数值越大,与实际情况相符。研究表明,所提出的对虾游动活跃性定量检测方法可准确快速获得游动活跃性指标,能高效地量化凡纳滨对虾在饲料盘上的活跃状态,对掌握养殖过程中凡纳滨对虾健康状况具有重要意义。Shrimp is rich in a variety of trace elements and vitamins,and has a substantial nutritional value,it is also be widely recognized as an important ingredient in high-end,well-known cuisine.Among them,the cultural production of Litopenaeus vannamei accounts for about 85%of the total production of shrimp culture,which is an important economic aquaculture object.The active state of L.vannamei reacts its health condition and behavioral situation.Surveying and identifying the activity of L.vannamei is helpful in finding abnormal behavior in aquaculture,to give early warning and take remedial methods promptly,lessen economic losses in aquaculture,and improve the yield and efficiency of aquaculture.Nowadays in the L.vannamei pond aquaculure process,aquaculture personnel often need to monitor the active swimming state of the shrimp by manually pulling the feed tray,then analyzing the overall environment of the aquaculture pond and formulating effective aquaculture breeding strategies.However,due to the complexity of the pond underwater environment,artificial observation experience is limited,so the method of manually observing the active state of L.vannamei has a lot of problems,such as inefficiency limited scope of application,low accuracy,poor real-time performance,high labor intensity and other problems.In order to solve these problems,propose a visual detection method for the activity of L.vannamei based on an improved YOLOv7 tiny network detection model and multi-objective association based on Euclidean distance to quantitatively study the swimming activity status of shrimp.Based on the YOLOv7 tiny network model,the standard convolution was replaced by Conv convolution,and a VoVGSCSPC module was built to replace the original lightweight aggregation module(ELAN-L).The MPDIoU loss function was used instead of the CIoU loss function to reduce the model capacity and improve the model detection accuracy.The position of shrimp in the image was determined by the visual detection results of improved YOLOv7-tiny model and th

关 键 词：凡纳滨对虾 游动活跃性 机器视觉检测 YOLOv7-tiny 池塘养殖 

分 类 号：S968.22[农业科学—水产养殖] TP391[农业科学—水产科学]