Research Of Fish Feeding Behavior Detection And Accurate Feeding Decision Method Based On Deep Leaning

In the process of aquaculture,feeding is a key link that affects the production cost and the growth and welfare of fish.Analyzing the feeding behavior of fish is an important basis to make accurate feeding decisions.At present,the baiting machine for feeding fish is still mainly based on manual experience,which leads to excessive or insufficient baiting.However,the identification of fish feeding behavior can effectively provide a basis for accurate feeding decision making.Due to the complex culture environment and the uncertainty of fish behavior,it is a challenge to accurately identify the feeding behavior of fish.In order to improve the accuracy of fish feeding behavior identification and optimize feeding strategy,an intelligent feeding decision-making method based on deep learning and fish feeding behavior analysis was proposed.Firstly,pre-processing operations such as contrast enhancement were carried out on fish feeding images to construct image datasets containing different lighting environments and feeding processes.Then,the improved YOLOv 5s algorithm was used to recognize the fish feeding image,and the key indicators that could quantify the feeding behavior,such as aggregation and swimming intensity,were extracted by combining Delaunay triangulation and Deep SORT algorithm.Finally,ANFIS feeding prediction and decision model based on feeding behavior,body weight and water temperature information was constructed to realize feeding demand during fish growth.It provides a parameter basis for the development of precision seed metering device.This research was carried out from the following aspects:Aiming at the problems of small number and large number of underwater fish targets and individual occlusion during the movement,a target recognition algorithm based on improved YOLOv 5s was proposed.By pruning the YOLOv 5s feature extraction network and deleting 20×20 related feature layers,the detection speed of the model for small targets was improved.By adding the attention mechanism module of CBAM,the recognition ability of the model to the occlusion information of underwater fish was enhanced.The model test results indicated that compared with the original YOLOv 5s model,the optimized YOLOv 5s model improved the detection accuracy P by 5.1% and the running speed by 22.1%.It showed that the YOLOv 5s model optimized in this paper had good effect on detection accuracy and speed,which provided an important basis for the aggregation of fish feeding behavior index and the quantitative analysis of swimming strength.Aiming at the accuracy of traditional machine vision technology in processing occluded image information,a method based on YOLOv 5s combined with Delaunay triangulation and depth sorting target tracking was proposed to quantitatively analyze the feeding behavior of fish and extract the key indicators that could quantify feeding behavior: aggregation and swimming strength.The experimental results showed that the degree of fish aggregation had the same changing law under different feeding desires.In the process of feeding,FIFFB quickly dropped from the initial value to the lowest level,and after maintaining a relatively stable time,it gradually recovered to the original level.Compared with high appetite fish,the recovery time of medium and low appetite fish was 15 s and 25 s earlier,respectively.When the fish’s appetite for food was high,medium and low,the swimming intensity rapidly increased from the initial value to the highest level,which was 17.73 cm/s,16.36 cm/s and 12.78 cm/s respectively,and then gradually returned to the initial level.This feeding law provided a theoretical basis for real-time feeding desire feedback and control strategy,and realizes accurate feeding of fish.Aiming at the problem that the current feeding control system can’t make accurate feeding control decisions according to the feeding desire of fish,an adaptive network fuzzy reasoning feeding decision system based on fish feeding behavior,biomass and weight information and water temperature was proposed.Taking the aggregation degree,water temperature and average weight as the inputs of the feeding decision-making model,the experimental data of feeding were fuzzified,the membership degree was selected,the feeding decision-making control rules were established,and the feeding amount was accurately controlled by model training and testing.The simulation and experimental results showed that the root mean square error RMSE and the average absolute error average error of this model were 0.78 and 0.19,respectively,and all the indexes were smaller than those of the traditional FIS prediction model.In terms of fish growth performance,the feed utilization rate FCR of ANFIS prediction model was about 12.90% higher than that of FIS model.At the same time,the relative error of body weight FMAE increased by 31.28%,which indicated that the fish specifications in ANFIS prediction model feeding mode were more unified.On the premise of improving feed utilization rate,this method had more advantages in fish growth specifications,which could effectively improve the economic benefits of producers and provide theoretical basis for the development of intelligent feeding equipment decision-making systems.