Marine Diatom Identification Research Based On Deep Learning

Diatoms are ubiquitous phytoplankton that represent the primary source of photosynthesis(and oxygen production)in the ocean,although many are freshwater inhabitants.Diatoms play an important role in ecological monitoring,Mineral monitoring,biopharmaceuticals and other fields.Therefore,the detection of diatoms is a meaningful subject.This paper designs a one-stage diatom detection network model,which optimizes the characteristics of diatoms from the dataset and network structure.The main work is as follows:1.In terms of improving the accuracy of diatom detection,random sampling fusion dataset is used to enhance the diatom dataset.By adding focus structure and residual structural design characteristics to keep diatom texture information.At the same time,the path Aggregation Network structure is integrated,so that the texture information of the diatom can directly enter the prediction network,then the detection network is less dependent on the shape characteristics of the diatom,and more texture features are learned.2.In dealing with the diatom stacking problem,stacking dataset image is used to simulate the stacking situation,and the proportion of diatom stacking images in the dataset is increased to 10%.It has been proven to effectively improve the recognition accuracy of stacked diatoms.Similarly,when facing the problem of different life cycles of the same species of diatoms,the random dataset sampling and splicing method is used to balance diatoms dataset.In addition,it combines batch normalization and spatial pyramid pooling to effectively improve the accuracy of small diatom recognition.3.Finally,splitting and combining images is used to directly detect the diatom pictures under the microscope with large pixels.This paper proposes an efficient prediction box merging algorithm to effectively solve the same target prediction box split problem.All prediction boxes can be processed at the time complexity of nlog(n)level.Finally,the training test is carried out through Coscinodiscusexcentricus and Hemidiscuscuneiformis Gr dataset.comparing with the current mainstream yolo series,SSD,faster-RCNN and other target detection networks.our diatom detection network score on the experimental platform is m AP(0.5)93.7,and the processing time is38 ms.It can also process diatom pictures under the microscope with a resolution of22577×13498 in 41 s,which is convenient for related researchers.