Deep Learning Based Fish Feeding Intensity Detection And Intelligent Feeding System

China is a large aquaculture country,but it is not a powerful aquaculture country.In the breeding process,the degree of automation and information is low,and the key feeding and other operations are still mainly controlled by time sequence,resulting in the waste of feed and the corresponding increase in labor costs.Therefore,efficient and accurate feeding system is particularly important.Previous studies have shown that accurate detection of feeding intensity of fish can help achieve accurate control of feeding amount during feeding process,thus controlling bait cost and improving production efficiency.To solve the above problems,this study took rainbow trout in factory recirculating aquaculture as the research object.Using the deep learning and machine vision technology,this paper proposes a real-time,high precision,lightweight 3D Res Net-Glo Re fish feeding intensity detection algorithm,and develops based on the fish feeding intensity classification intelligent feeding system.It can effectively overcome the challenges brought by uneven illumination and fish movement to image processing,and help improve the automation and intelligence level in the process of bait casting.The specific research contents are as follows:(1)A real-time,high-precision and lightweight 3D ResNet-GloRe fish feeding intensity detection algorithm is proposed,which can effectively capture the characteristics of splash and texture caused by fish movement,and then detect the feeding intensity of fish in the video stream in real time.First,the lightweight Glo Re module was 3D expanded,and the Residual block in 3D Res Net network was changed to 3D Glo Re module to form 3D Res Net-Glo Re network.3D Glo Re module realizes relational reasoning through graph convolution in interactive space,which improves the discrimination accuracy of network.At the same time,on the premise of ensuring the accuracy,the video is processed with frame extraction,which can reduce the number of model parameters and the amount of calculation,and speed up the training and recognition of the network.The experimental results show that the accuracy of the proposed algorithm reaches 92.68%,which is 4.88% higher than the classical 3D Res Net algorithm.Meanwhile,Parameters decreased by 46.08% and GFLOPs by 44.10%,which can predict 17 fish feeding intensity per second.(2)Developed a real-time detection and training system for fish feeding intensity,and developed an intelligent feeding system based on feeding intensity classification.Combined with 3D Res Net-Glo Re algorithm and Py QT5 technology,the real-time detection and training system of fish feeding intensity realized the functions of model training and real-time monitoring of fish feeding intensity.The intelligent feeding system includes feed box,feeding device,throwing device and intelligent speed control module.The speed control module sets the program of four speed and controls the speed of the feeding device,so that the four-speed of the feeding device corresponds to four grades of feeding intensity of fish.Through the real-time detection system to accurately analyze the video stream of fish feeding,the current feeding intensity label of fish is obtained,and the corresponding feeding instruction is given.Then,based on feeding instruction,intelligent feeding system is controlled to realize accurate and real-time graded feeding.Finally,the bait casting speed can be adjusted automatically according to fish demand,so as to achieve the purpose of fine feeding and cost reduction.