Researches Of Classification Algorithm Applied On High-Resolution Microscope Images

Within recent decades,nanotechnology has developed rapidly and has made remarkable breakthroughs in the agricultural field.Specially,some progress has been made in preparation of nanomaterials and creation of new agricultural investment products.The cross integration of nanotechnology and traditional areas has greatly promoted the technological revolution of the eco-agriculture industry.As one of the main branches of nanotechnology,nano-microscopic technology has received extensive attention from researchers.Among them,quantitative determination of the atomic structure of nanomaterials is an important part of the nanometer microscopy technology,which is beneficial for people to study the potential properties of nanomaterial microstructure.The atomic structure includes the atomic position and its element type,the current atomic position can be obtained by measuring the intensity center,while the element type can be quantified by the magnitude of image intensity.The traditional method uses the numerical distribution of the intensity to determine the elements.Although this approach can achieve a good classification,it is not efficient and subjective.Therefore,it is of great significance to use the machine learning algorithms with good self-learning ability to solve the problem of atomic column composition classification.In this paper,the classification algorithms in machine learning are applied to the study of high-resolution microscopy images,for solving the problem of identification and doping analysis of atom columns.The main contents are as follows:（1）Brief introduction to the theories of machine learning.This section includes data processing methods and common classification algorithms,which lead to the concept of the ensemble learning.Then,the ensemble learning algorithm is described in detail,which is based on the generation of base learner,integration strategy and common algorithms.Aiming at the atomic recognition and classification problem proposed in this paper,we introduce the basic principle and algorithm derivation of the extreme gradient lifting and the mean shift clustering algorithm.（2）The dataset acquisition and feature engineering.In this part,we use 3D block matching algorithm to filter images,multi-ellipse fitting method to determine atomic position and integrate intensities of pixels within a certain range to obtain average intensity of each atomic column.In the feature engineering of the dataset,the relative information of the neighboring atoms is taken as the feature of the target atom based on the structural periodicity,including distance,angle and intensity value.The importance of features is sorted by the extreme gradient lifting algorithm and the construction of sample collection is completed by selecting the appropriate feature.（3）Application and comparison of different algorithms in element classification.We study the experimental images of Nb_0.8CoSb and MoS_2（1-x）Se_2x.Firstly,the extreme gradient lifting algorithm is used to identify and classify the atoms of these two samples,and the prediction accuracy of model is studied very well.Secondly,its performance is evaluated by comparing the algorithms of the K-nearest neighbors and the support-vector-machine algorithms and studying the classification prediction of the model after the reduction of training set.（4）The doping analysis for the experimental image.The mean-shift clustering algorithm is applied to analyze the doping atoms at the S sites in the MoS_2（1-x）Se_2x,which may be S₂,SSe or Se₂ only according to the atomic-column intensity.Thus,the doping ratio can be calculated.The experimental results in this paper are in line with the expected goals,which can effectively solve the problem of atomic-column composition classification and doping analysis.The algorithms are simple to implement with high training efficiency,able to quickly obtain classification results,and greatly reduces the workload of manual operations.And this work provides a certain theoretical and experimental basis for the post-processing of large-scale atomic classification problems.Additionally,a related software has been developed with a friendly user interface,which is supposed to facilitate the analysis work for other researchers in microscopy field.