Research And Implementation Of Temporal Detection Method For Fish Feeding Action

Feeding is one of the important aquaculture operations in the aquaculture process.The traditional feeding method uses baiting machines to feed regularly and quantitatively,which can easily lead to under-and overfeeding.Changes in feeding behavior of fish can effectively reflect their feeding levels,which can be used to precisely control the feeding amount and help to solve the above problems.Based on this,this study takes golden trout in factory recirculating water aquaculture as the research object,uses deep learning,proposes a high-precision BMN(Boundary Matching Network)based fish state detection network BMN-Fish,and developed a smart fishery precision control system,which effectively solved the problem of low accuracy of previous algorithms in switching the critical point of fish feeding action,as follows:(1)A dataset of fish feeding action temporal detection task was constructed.The experimental environment was built according to the characteristics of the breeding environment,and the video data containing the complete feeding actions of the fish required by the model were collected,and the video times all ranged from 20 s to 200 s,with a total of 100 videos.Since the fish action categories are relatively single,which leads to easy overfitting in network training,this study merges the collected fish video data with the public dataset Activity Net-1.3 to form a large dataset with 201 action categories.(2)An improved BMN-based fish state detection network BMN-Fish is proposed.by modifying the base module(Base)into a basic residual module(BRM),the residual structure is introduced into the convolutional neural network,which protects the integrity of information by passing the input information around to the output,and can focus deeply on the region of interest in the feature map to obtain more effective feature information;then the efficient channel attention module(ECANet)is added to enhance the global information extraction capability by using the attention mechanism,which in turn effectively expands the perceptual field of temporal dimensional features and improves the network performance.The experimental results show that the AUC of BMN-Fish reaches 93.32%,which is 2.17% higher than that of BMN.At the same time,the average recall rate(AR)of the nomination number(AN)is 100,namely AR@100 There was an increase of 1.95%.The improved network can detect fish feeding action more accurately and can provide algorithmic support for developing accurate feeding systems.(3)The intelligent fishery precision control system is developed and integrated with the fish state detection network BMN-Fish proposed in this paper,which can detect the video of fish feeding action,get the start and end time of fish feeding action,support data pre-processing and model training;and provide a visualized and interactive multiparameter real-time online aquaculture environment monitoring and precision control system for fishermen and researchers.The system can continuously realize the operation of information collection,information processing,information display,and precise control 24 hours a day,thus helping fishermen and researchers in the fishery field to monitor and analyze the feeding behavior of fish.