Reconstraction And Recognition Of Dark-field Atom Images Under Electron Microscope

In the field of materials science,the images quality of dark-field atom images is an important visualization basis the scientific research results of material scientists.When material scientists need to prove the material,they have developed has excellent uniformity,the arrangement and distribution of atoms on the surface of the material becomes an indispensable condition,and only a sufficiently clear image of the dark-field atom and accurate atom center coordinates and radius can be used as sufficient evidence for the dark-field atom arrangement and distribution for materials scientists.With the development of deep learning technology,image processing and recognition technology has gradually been applied to many fields such as medical images,satellite maps and security fields in recent years.The dark field atom images acquired by scanning transmission electron microscopy are difficult to meet the research needs of researchers in the field of materials science due to problems such as experimental environment,material quality and system image quality.However,traditional image processing methods cannot effectively improve the signal-to-noise ratio of dark-field atom images,improve image quality and identify and analyze image content.Therefore,using deep learning technology to reconstruct and analyze dark-field atom images has become a solution to solve the problem of images quality.This paper mainly solves two problems,namely,reconstruction and recognition of dark field atom images.In order to improve the image quality of dark-field atom images,the paper proposes a dark-field atom image reconstruction model based on the generative adversarial network.Firstly,the EMD-ZGP loss function is designed to solve the difficulty of training,which improved the stability and the training effect of the model,and accelerated model convergence.Secondly,the DSN-U-Net network is redesigned as a generator based on the U-Net at the generator side.Combining the methods of hollow convolution and spectral normalization to improve the quality of the generator,it makes atoms' shape more regular and the brightness clearer and less noise in dark-field atom images.Thirdly,at the discriminator side,the appropriate training scheme and optimizer are selected based on the loss function and the characteristics of the generator,which further improves the reconstruction quality of the dark-field atom image.Aiming at the recognition problem of dark-field atom image,this paper firstly adopts the neural network model based on Faster R-CNN.Secondly,it combines the image reconstruction model when constructing the enhanced datasets,which realizes the accurate recognition of dark-field atoms and the accurate calculation of central coordinates and atom radius.At last,in order to deal with the recognition task of multi-size dark-field atom images,the recognition module is designed and expanded to realize the accurate recognition of dark-area atom images of different sizes.Through the above several solutions,the problem that cannot be solved by traditional methods in the field of material science has been solved,which has provided strong support for the research work of reconstruction and recognition of dark field atom images under electron microscope.