The Observation Method Of Phytoplankton Fluorescence Image Based On Deep Learning

Phytoplankton are one of the most important primary producers in marine ecosystems and important indicators for marine ecological disasters such as harmful algal blooms.Achieving rapid quantitative analysis of marine phytoplankton is an urgent need for marine ecological scientific research and environmental protection.Although some observation methods such as light microscopy,remote sensing,and imaging flow cytometry are widely used in marine environment monitoring,these methods still have limitations including very low detection throughput,poor observation accuracy and high-cost,which are not able to efficiently,quickly and accurately observe phytoplankton.Therefore,in this thesis,based on the Fluo Sieve^?fluorescence imaging flow cytometer,we develop image classification and object detection algorithms based on deep learning and design an embedded system for intelligent image acquisition and analysis of marine phytoplankton to achieve high-throughput end-to-end image processing.It provides a powerful solution for marine environmental monitoring and protection.The major research achievements in this thesis include the following three parts.1.Construction of a phytoplankton fluorescence image dataset.The Fluo Sieve^?instrument combined with a set of the image construction process based on the idea of active learning is used to carry out image acquisition and data cleaning for common phytoplankton in the South and East China Sea.A large-scale marine phytoplankton fluorescence image dataset containing a total of 418,497 images in 32 different categories is constructed.2.Development of intelligent image processing algorithms for phytoplankton.Using the large-scale phytoplankton fluorescence image dataset,we develop the phytoplankton image classification algorithm based on Res Net-18 network and the phytoplankton object detection algorithm based on YOLOX network,both of which are optimized and accelerated to meet the requirements of embedded deployment and real-time processing.3.Embedded system implementation and application of intelligent phytoplankton image acquisition and analysis.In this thesis,from the perspective of instrument system architecture,we implement an embedded system integrating pump valve fluid manipulation subsystem,laser control subsystem,and real-time image acquisition and intelligent analysis subsystem.The three subsystems work together to achieve high throughput and rapid intelligent processing and analysis of marine phytoplankton samples.