Neural Network-based STM Image Segmentation Of Nanostructures

For the segmentation of the nanostructure STM image,the purpose is to identify the nanostructure area from the complex background,and then assist in the analysis of its physical characteristic parameters.In order to improve the accuracy of nanostructure image segmentation,a neural network-based STM images segmentation method is adopted.Take nanostructures such as quantum dots as regions of interest to extend the image segmentation experiments.Which mainly involves image preprocessing,segmentation,physical feature parameter recognition and so on.The preprocessing experiment mainly includes the processes of grayscale,noise reduction,decontamination and data enhancement.First,according to the image characteristics,an image grayscale algorithm based on adaptive weights is designed to reduce the data dimension.Then,the mask method is used to optimize the image mean filter to achieve the purpose of eliminating pollution.Finally,a Ga As(001)substrate,quantum dot and quantum ring material label data set containing 900 images is produced by manual marking,and the data is enhanced by graphic transformation.In the segmentation experiment,firstly,neural network algorithms are used such as FCN,Seg Net,U-Net to achieve the segmentation of the region of interest.At the same time,the contradiction of the original network pixel classification and positioning is coordinated and the segmentation quality is improved by introducing an intermediate convolutional layer in U-Net.The improved U-Net's average Dice coefficient,accuracy rate,accuracy rate,recall rate and IOU value in cross-validation reached 93.99%,92.22%,96.21%,92.05%and 87.94% respectively,which are all higher than the original network.In the identification of physical characteristic parameters,the size,roughness,firstly,quantity and density of the Ga As(001)substrate platform are identified by the connected domain statistical method.Secondly,the smallest circumscribed rectangle marking method is used to extract the aspect ratio of the nanostructure.Finally,an algorithm based on standard uniform distribution matching is designed to calculate the uniformity of nanostructure distribution.In summary,in order to improve the quality of nanostructure STM images segmentation,this paper uses an improved neural network method to achieve image segmentation.The segmentation result of the algorithm used effectively improves the identification accuracy of the physical characteristic parameters of the nanostructure from the verification.