Deep Learning Based On Detection And Segmentation Of Nanoparticle Structures

Since nanoparticles have a wide range of applications in many scientific fields such as physics,chemistry,electronics,optics,biomedicine,and materials science,it has received a lot of attention from researchers.The type of properties depends on several characteristics,some of which are related to the particle structure,so the correct characterization of nanoparticles is crucial.For structural characterization of particles and processing of experimental images of micro and nano structures,the traditional techniques are electron microscopy and image processing techniques.However,the process of electron microscopy is tedious,repetitive and slow;the image processing techniques are demanding for the images,and usually,the nanoparticle images are very noisy and have problems such as uneven brightness distribution and blurred edges,and their algorithms all have to determine the corresponding parameters for different images to obtain better results.Therefore,it is easy to be affected by the noise in the image,and the processing results are difficult to be controlled.In recent years,deep learning has shown significant advantages in image processing.In this paper,we identify and segment the structure of nanoparticles based on deep learning techniques,and improve on the original model,and the constructed models all achieve very good results.The main research contents of this paper are as follows.(1)Image datasets for nanoparticle structure recognition and segmentation were constructed,respectively.The ordered dataset obtained by BOIKO using electron microscopy was annotated,and the datasets for recognition and segmentation were constructed separately,and the special cases in the structures were specially labeled and processed,and finally the datasets were expanded by data augmentation,which laid the foundation for carrying out the deep learning-based recognition and segmentation models,respectively.(2)A study on the recognition of nanoparticle structures based on Faster R-CNN.The Faster R-CNN network structure was used to identify the nanoparticle structures,and it was able to accurately identify the circular structures in the graph with 98.663%Mean Average Precision(MAP).(3)Segmentation study and improvement of nanoparticle structures based on UNet.Firstly,using dozens of segmentation models,observational studies were conducted to determine the segmentation study of the structure using UNet,and then UNet was improved by incorporating the improved Dense Block module Lightweight Dense,ASPP module and Spa module,and the DSPUNet model was proposed with three states,DSPUNet-a,DSPUNet-b and DSPUNet-c,and the parameters are optimized,and finally the optimal Mean Intersection over Union(MIo U)ratio of which reaches 0.964.In summary,with the help of deep learning methods,we have successfully established deep learning recognition and segmentation models for the identification and segmentation of nanoparticle structures,respectively,and improved the models to propose models applicable to the study of nanoparticle structures.At the same time,we produced a nanoparticle structure recognition and segmentation dataset to pave the way for subsequent deep learning-based studies of nanoparticle structures.This work will contribute to the further research and development of nanoparticles.