Research On Shrimp Keypoint Detection And Individual Recognition Algorithm Based On Light-weight Deep Learning

Phenotypic data is an important observation index in the process of shrimp breeding and optimization.Traditional methods of measuring phenotypic data mainly rely on manual measurement,which not only has high costs regarding manpower,but also has low efficiency and limited phenotypic data.Therefore,how to efficiently obtain the phenotypic data of shrimp has been perplexing the shrimp breeding community.Further,the process of shrimp breeding and optimization required to distinguish each shrimp family,now only artificial way to classify the family,which severely limits the process of shrimp breeding optimization.Recently,deep learning algorithms have played a huge role in various industrial fields with their excellent performance and robustness in dealing with complex scenarios.Therefore,this thesis takes the Pacific white shrimp as the research object,and is committed to researching shrimp body key point detection and individual identification algorithms that can balance calculation consumption and model accuracy,and realize a stable and easy-to-use intelligent shrimp data collection system.The main work and innovation results of this thesis are as follows:1.On the basis of the existing key point detection network model,based on the optimization and improvement of the hourglass network,this thesis proposes a lightweight shrimp key point detection network that balances model accuracy and operating efficiency.This thesis redesigned the structure of the bottleneck module and the up-sampling module in the hourglass network.The final network model size is only20 MB,the calculations(GFLOPs)are only one twentieth of the original network,and the accuracy of the model is only about 3% lower than the original model.2.Based on the sensitivity of the shrimp body,a method of marking shrimp tails based on visually filled silica gel dyes was designed,and based on the lightweight key point detection backbone network,a color point classification module was added to achieve accurate and fast shrimp.The individual tag recognition network has an individual recognition accuracy of 95%.3.Based on the above algorithm model,a client/server model is used to design and implement an intelligent shrimp data collection system.And based on the needs of shrimp breeding and aquaculture researchers,the functions of user login,user management,equipment management,data query and data collection of the intelligent shrimp data collection system have been realized.