Research On Individual Recognition Of Takifugu Bimaculatus Based On Deep Learning

Puffer fish is a precious fish with delicate flavor but highly toxic,presenting great market prospects.However,given that some puffer fish are highly toxic and the currently used traceability technologies separate the tag from the body of puffer fish,such as QR code and electronic tag,which does not fundamentally solve the problem of traceability.Therefore,this thesis takes takifugu bimaculatus,a cultured species of pufferfish in Fujian Province,as the research object(unless otherwise specified,the puffer fish described in this thesis are takifugu bimaculatus)and proposes a method of using biological image features to trace the source of puffer fish.This method divides the recognition process into four steps,including the image extraction,image alignment,feature extraction,and feature matching of puffer fish.First of all,for the recognition of puffer fish,the thesis builds a puffer fish dataset for image segmentation and a dataset for extracting the features of puffer fish.The former includes a total of 1267 pictures of 30 puffer fish while the latter includes a total of 55837 pictures of 120 puffer fish.Secondly,regarding the image extraction of puffer fish,this study explores the extraction methods based on image segmentation.After analyzing and comparing multiple image segmentation models through experiments,the thesis presents DeepLab V3+,a model with a better segmentation effect.Based on image segmentation,the study proposes an alignment method based on the mask,which utilizes the mask shape of puffer fish for alignment.Experiments show that this alignment method is simple and efficient,capable of aligning puffer fish in any direction.Then,the thesis studies the methods of feature extraction of puffer fish.After specific experiments,the study found that traditional algorithms for feature extraction,such as LBP,perform relatively low accuracy,which is,generally below 50%.For algorithms based on convolutional neural networks,this thesis explores nine different models combing basic network and loss function.The accuracy of these nine models in the feature extraction dataset is generally above 99%.Meanwhile,given that the texture of puffer fish is characterized by a long shape,the study introduces a rectangular convolution kernel in the neural network algorithm with high accuracy,and its effectiveness is verified through experiments.Finally,this thesis conducts an overall test analysis of the complete recognition model,verifying its performance in identifying known individuals and unknown individuals respectively.This model has a simple structure,convenient in testing,and fast in recognizing puffer fish,and can be used in the traceability of food safety as well as the protection and tracking of marine resources,which is of great social and economic significance.