Study On Fish Behavior Classification Based On Deep Learning

In the context of smart fishery,this paper combines deep learning with the detection of fish behavior to identify and classify the behavior state of fish school,aiming at achieving accurate control of the breeding environment,understanding the growth status of fish and determining the health of fish.Under the influence of different environmental factors,the behavior state of fish school has different changes,including movement speed,reproduction ability,learning and memory ability,avoidance behavior,social behavior,etc.These changes are not only reflected in the spatial distribution of the group,but also reflected in the time distribution of the group and individual,like the movement speed.First of all,this paper carried out experiments on the behavior state of zebrafish school.By collecting data in the laboratory environment,a sample database of the behavior state of zebrafish school was established.Typical neural networks LeNet,VGGNet and ResNet were selected to extract sample characteristics through single-channel neural network,and the single-channel neural network experiment was completed.Secondly,simulated biological vision system to build neural network model of the structure of the two-channel,fish behavior states information can be divided into two parts: time and space which went through space network and time network respectively.After two kinds of channel handle sample space information and temporal information,the network carries on information fusion,completes the two-channel neural network experiment.Finally,the residual network(ResNet)was selected as the single channel network of the two-channel neural network through experimental comparison,the fish-school behavior detection algorithm model based on the deep neural network was constructed.In the network,a series of strategies such as migration learning,batch normalization processing and dropout were adopted to solve the problem of training overfitting,so as to enhance the learning and generalization ability of the network.The results verified the effectiveness of the two-channel network used in this project in the task of fish behavior classification.