Research On Underwater Fish Individual Recognition Method Based On Deep Vision Perception

In fisheries resources exploitation and management,fish individual recognition is very significant.Precision farming is one of the main applications of fish individual recognition technology in the field of fish farming.It can track each fish in real time,conveniently grasp the growth condition of each fish,and scientifically analyze the growth data and feeding amount of each individual in different periods.Both reduce the waste of resources and improve the efficiency of farming.Fish individual recognition technology can be used to detect epidemics in fish,understand their behavioral habits,and deal with abnormal behaviors promptly.Effectively reduce the time to save the disease and develop a reasonable protection and management plan.The recognition of individual fish can accurately predict their length,weight,and maturity,helping aquaculture companies to monitor the growth of individual fish in real-time and to classify individual fish in the aquaculture area promptly.It can effectively avoid the transmission of epidemic diseases,and at the same time,it can deliver the mature fish to the market in time to avoid over-farming.Currently,in the progress of research on fish individual recognition,there has been a shift from traditional methods such as pattern and shape to machine vision recognition.Although fish individual recognition techniques using computer vision can substantially improve recognition accuracy,in the actual underwater ambiguous environment,due to uneven illumination,more background noise,occlusion,and inconspicuous intra-class features.The method based on traditional machine vision has the problem of difficult and slow recognition.Therefore,to solve these problems,this paper proposed a fish individual recognition algorithm based on detail feature enhancement and a fish individual recognition algorithm incorporating deformable convolution and channel groupings,respectively.The main research contents are as follows:(1)In the practical application situations of fish individual recognition,due to more noise,an angular tilt of the fish body,and the insignificant intra-class feature differences,resulting in convolutional neural network feature extraction capacity reduced,affecting the problem of recognition accuracy.A fish individual recognition algorithm based on detail feature enhancement(Fish Net-v1)is proposed.To solve the problem of possible angular tilt of fish pictures,key points detection is added in the target detection stage to process the data.And use the optimized Mobile Net-v1 feature extraction network,combined with the feature inversion module and the fusion module to learn the fish features more comprehensively.The final recognition result is obtained by feature comparison.(2)In the further study of fish individual recognition,we found that the large-scale variation of fish pictures and the inability to completely remove the background information outside the fish body still affect the recognition performance of the model.Therefore,to further improve the accuracy of fish individual recognition and at the same time reduce the number of network parameters to improve the network training efficiency.A fish individual recognition algorithm incorporating deformable convolution and channel groupings(Fish Net-v2)is proposed.Further,optimize the target detection module to improve the network’s multi-scale target detection accuracy.Meanwhile,the depth separable convolution is replaced by the deformable convolution in the recognition module,it makes the perceptual field of the convolution kernel closer to the morphology of the fish body and removes the redundant background information.And grouping and separating the convolution of feature maps according to channels.Use a weighted feature layer to replace the global average pooling layer in the network structure,combined with the funnel activation function to improve the model’s ability to capture spatial information correlation.Finally,it forms the feature information that can be used to recognize the ID,thus realizing the improvement of the accuracy of fish individual recognition while reducing the computational effort.