Researches On Imaging Conditions Of Electron Microscope Based On Neural Network

With the rapid development of computer technology and the arrival of the era of Big Data,breakthroughs have been achieved in a large amount of fields with deep learning technology used.Among them,the convolutional neural network(CNN)is particularly prominent in the field of image processing because of its excellent characteristics such as local receptive field,weight sharing and down-sampling.Thus,we combine it with the electron microscope(EM)image to determine the imaging parameters,which is one of the most basic problems in the EM field.In the experiment,the lens aberrations of EM are measured from the diffraction images of the amorphous region while the interesting region is the crystalline lattice.Thus,the simulation image of the crystalline lattice simulated with the aberration parameters directly provided by the EM does not agree with the acquired experimental image.In order to measure the correct aberration parameters of the experimental image,we can generate a large number of simulation images and find out the one which is the most consistent with the experimental image,through which,we can accurately obtain the imaging parameters.In this paper,the CNN is used to automatically recognize this simulation image,which is likely to be consistent to the experimental image.The main contents are as follows:(1)A neural network is trained by a suitable data set.Based on the parameters of the image provided by the microscope,a large amount of simulation images are generated by changing the parameters in a certain range.Since the totally 21 imaging parameters should be determined,in order to reduce the number of random numbers in each parameter range,the method of reducing the data magnitude and the precision is adopted.(2)The conventional parameters are refined by using the method of parameter transfer in transfer learning,since it is greatly hard to retrain a set of completely new parameters and the conventional parameters of neural networks are less adaptable to the EM images.Also,an unsupervised clustering is used to reduce the simulation image data scale because existing equipment cannot meet the large data scale.(3)The CNN method is used to recognize the EM image.Using the clustered data set and the parameters after parameter transfer,we modify the loss function of the network to train a model used for EM image recognition.(4)The feasibility and validity of the proposed scheme are verified by testing simulation images.And the experimental images of graphene are employed to determine their imaging parameters.The experimental results are in accordance with the values achieved by manual measurement,which requires several days to get.(5)Preprocessing of experiment image.We have proposed a new method of the multiellipse fitting method to accurately determine the periodic positions on image.According to these positions,some intercepted small regions are averaged to reduce the image noise.Through this,the image detail affected by the imaging parameters is very clear and good for the image recognization by using the CNN.