| As a valuable protected freshwater fish species,Varicorhinus Macrolepis has high medical and nutritional value,but it grows slowly in the natural environment and has certain requirements for the living environment.Feeding is a necessary part of large-scale breeding of Varicorhinus Macrolepis.The machine vision is used to analyze and compare the characteristic information before and after feeding,so as to provide reference for the formulation of intelligent feeding strategy.The main research contents and conclusions of this paper are as follows :(1)A target detection method of Varicorhinus Macrolepis based on convolutional neural network YOLOv5 is proposed.Firstly,2000 samples are obtained by collecting image samples of Varicorhinus Macrolepis and data enhancement.The Varicorhinus Macrolepis frame is calibrated and the data set is prepared.Secondly,the four versions of YOLOv5,YOLOv5 s,YOLOv5m,YOLOv5 l and YOLOv5 x,which exist in YOLOv5,are uniformly used Mosaic data enhancement,cosine annealing attenuation learning rate and adjustment hyperparameters to carry out model training and selection test.Based on the comprehensive consideration of recognition accuracy,model complexity and resource utilization,YOLOv5 s model is determined as the best application model of this study.Finally,further lightweight improvement is carried out for the selected model.The improved test results show that the average accuracy value m AP,accuracy rate P and recall rate R of the lightweight model exceed the original model by 0.3 %,0.1 % and 0.3 % respectively when the parameters are reduced by 30 %,and the performance is equivalent.(2)Based on the lightweight improved model,the research on the observation method of feeding behavior of Varicorhinus Macrolepis is proposed.The observation is carried out from three aspects : swimming speed,clustering level and flow statistics.Firstly,the camera calibration experiment is carried out to test the image distortion.Then,the lightweight target detection model is combined with the Deep Sort tracking algorithm to realize the individual swimming speed reasoning of Varicorhinus Macrolepis.Secondly,the target detection module is combined with the DBSCAN clustering algorithm.The detection module is responsible for the extraction of target coordinate information,and the clustering algorithm is used to cluster and divide it.Then the target detection module realizes the aggregation and dispersion state recognition of Varicorhinus Macrolepis according to the division results.Finally,the module design of flow statistics is made.Based on the combination of lightweight model and tracking algorithm,the counting and timing modules are introduced to complete the flow statistics of Varicorhinus Macrolepis.The accuracy is more than 92 % by manual inspection.(3)The feeding observation experiment was carried out on the three established behaviors of Varicorhinus Macrolepis,including the time series observation experiment of feeding behavior of fish school in the same month,the comparison experiment of observation results with and without feeding arrangement of fish school in the same month,and the longitudinal ratio analysis of observation results in three different months.The statistics of swimming speed and flow rate showed that the activity level of Varicorhinus Macrolepis increased continuously near the feeding time,and decreased gradually with time after feeding.The control experiment with and without feeding showed that the activity level of Varicorhinus Macrolepis decreased more obviously than that without feeding during a period of time after normal feeding.Among the three different months of June,September and October,the activity level of Varicorhinus Macrolepis was the highest in June,followed by September and the lowest in October.The results of cluster level statistical test showed that feeding had no significant effect on the cluster state of Varicorhinus Macrolepis before and after feeding,and there was no significant difference in cluster level between different months. |
PDF Full Text Request