Research On Estimation Of Fish Density And Feeding State Based On Machine Vision

Smart fishery is a new mode of fishery development driven by modern information technology such as big data,IOT and artificial intelligence.It is an important way of supply side structural reform of aquaculture industry,involving the application and demand of aquaculture environment monitoring,biological environment monitoring and biological state monitoring.The population density estimation and feeding status quantification of fish were studied from the two hot points of production and health status.In this paper,the method of density estimation is used to complete the fish target count,and the number of targets is indirectly obtained by estimating the fish density.The feeding behavior of fish reflects the health status of fish.The performance of fish is different during the feeding period.Evaluating the strength of their feeding behavior is helpful to realize the digitization and informatization of fish health status judgment.The main research results are as follows:1.Create the density estimation dataset of fish.Firstly,the image dataset is constructed by collecting the fish image.Due to the species,size and density of fish targets in the dataset,the image area occupied by fish targets varies greatly,and the target size cannot be measured by the distance between targets.Because it is impossible to generate the true value of density by using the common Gaussian filtering method,the point graph is directly used as the true value of density.2.Design a fish density estimation network FDENet(Fish Density Estimation Network),which combines the advantages of skip connection,image pyramid input and feature perception to improve FDENet’s robustness of target size and the performance of density estimation.At the same time,the common loss function for estimated density LMseis not suitable for point graph density estimation dataset,and Lcount+Mse is the training loss function of FDENet.Finally,comparison experiment between different models,loss function comparison experiment and network structure ablation experiment are designed to verify the rationality and effectiveness of the proposed density estimation method.3.In order to solve the problem that the quantitative function of fish feeding intensity based on direct analysis adopts the single characteristics of fish swimming speed,which leads to the inaccurate reflection of feeding intensity,a quantitative function SLFI(Speed-Location Feeding Intensity)integrating fish swimming speed and spatial distribution is proposed.In order to solve the problem that the quantitative function coefficient of feeding intensity in indirect analysis is unreasonable,which leads to the inaccurate reflection of feeding intensity,the feeding intensity function Ea is proposed.Finally,the accuracy of the proposed method is verified by the fish feeding video.