Super-resolution Field Emission Scanning Electron Microscope Based On Deep Neural Network

The field emission scanning electron microscope is a complex scanning electron microscope with ultra-high-resolution image scanning,instant printing,and output storage capabilities.It has been widely introduced by universities and enterprises to analyze material's micro-morphology,cell organization,chemical composition,etc.However,there are still two problems in the use of the field emission scanning electron microscope.On the one hand,because of the research and development cycle of the core components is quite long,the imaging magnification cannot be further improved,that is,a higher resolution image cannot be obtained.On the other hand,the acquired images cannot obtain large field of view and high resolution at the same time.This contradiction between field of view and resolution is irreconcilable.These problems have limited the use efficiency and application scenarios of the field emission scanning electron microscopes to a certain extent.This paper studies super-resolution field emission scanning electron microscope based on deep neural networks.Using image super-resolution technology,it can effectively improve the resolution of the acquired images without changing the physical structure of the field emission scanning electron microscope.And it can also resolve the constraint between field of view and resolution,as long as taking the low-resolution image with large field of view as an input,a high-resolution image with large field of view can be quickly output.This paper proposed the super-resolution algorithm AARN based on residual learning.The network is composed of stacked asymmetric activation residual block AARB,the new upsampling model and a global residual connection.It can extract deeper image features and enrich texture details of the reconstructed super-resolution image.Validating the image quality evaluation indicators on the DIV2K dataset,the PSNR value is 36.82dB,and the SSIM value is 0.9719,which is significantly better than traditional mathematical methods and classic super-resolution algotithms.At the same time,the paper also proposed the the super-resolution algorithm AARN based on generative adversarial mechanism.The generator is developed from AARN network and the discriminator uses a convolution structure that is gradually constricted and thickened.The network model is trained with new loss function.Validating on the DIV2K dataset,the PSNR value is 35.67dB and the SSIM value is 0.9606,which is slightly lower than the AARN algorithm.However,the image reconstructed by the AARGAN algorithm is more real and natural,and its high-frequency information is more abundant.This paper will help to improve the imaging quality of the field emission scanning electron microscope,and has important significance in various application fields such as materials science,biology,medicine,metallurgy,geological exploration,criminal investigation and gem identification.