Basic Research Of Nano-structure Microscopic Vision Based On Support Vector Machine And Deep Learning

Microscopic vision,especially related to the identification,segmentation and parameter analysis of one-dimensional nanomaterials,is an important research field,which is of great significance for the establishment of the relationship between nanoscale materials and macroelectronic devices.Nanowires in nanomaterials have an important influence on photosensors,FET devices,etc.,and nanosheets have an impact on the efficiency of solar cells.Therefore,research in this field can promote the understanding of nanoscale materials.In this paper,the two materials are processed separately,and an effective algorithm for segmentation and recognition of these two materials is proposed.This paper first analyzes the characteristics of one-dimensional nanomaterial image of scanning electron microscope and the deficiencies brought by image recognition,segmentation and parameter extraction under this condition.Since the imaging principle of scanning electron microscope is based on secondary electrons,scanning electron microscope imaging The gray scale is not uniform,the background gray scale is not uniform,and the gray scale distribution of the nanowires is not uniform,which makes the segmentation recognition difficult.On this basis,two target objects of nanowires and nanosheets were studied.For nanowires,this paper analyzes the problems that nanowire segmentation needs to solve and the distribution characteristics of images.A Bayesian classifier based on optical channel hypothesis is proposed for image segmentation algorithm.The method in this paper can effectively solve the problem of uneven distribution of background gray scale and uneven gray scale of foreground nanowires.Experiments show that the proposed algorithm has better image segmentation effect than similar algorithms.At the same time,this paper further analyzes the problem of parameter extraction of SEM one-dimensional nanowires.Since the independent single nanowires,nanoparticles,and intersecting nanowires exist at the same time,the parameters for directly acquiring the nanowires are problematic,and the parameters of the nanowires cannot be directly obtained;the simultaneous presence of the nanowires and the nanoparticles may interfere with the segmentation.At the same time,the crossed nanowires will bring great difficulty to the parameter determination.In this paper,the multi-classification support vector machine is used to discuss the three objects mentioned above.Firstly,the characteristics of independent single nanowires are obtained.Later,the cross nanowires were discussed in detail,and the modified Hough transform strategy proposed in this paper was used to measure the boundaries of the nanowires in the cross nanowires.Experiments in image data samples show that the algorithm can effectively extract the parameters of nanowires.In terms of nanosheet structure,a large number of nanosheet structures are stacked together,and other objects are required to be small target objects,so that the general recognition algorithm cannot function.In this paper,according to the uniqueness of the nano-sheet structure,a deep learning framework combining the global information of the image and the local information of the image is designed,which overcomes the shortcomings of the small sample and the insufficient detail of the segmented image,and realizes the target in the SEM image of the nano-sheet structure.Recognition and segmentation of objects.The experimental results show that the deep learning network proposed in this paper can train well on small sample data,and the obtained image segmentation results are better than other deep learning networks that only use global information.