Recognition Of Nanocomposites Agglomeration In Scanning Electron Microscope Image With Semantic Segmentation Algorithm

Nano-dielectric is a new type of dielectric with the potential for performance editing.It can obtain excellent dielectric properties by homogenously dispersing nanoparticles in a polymer matrix,and is considered to be the most potential dielectric material in the future.However,nanodielectrics fail to achieve the expected performance in actual applications because of agglomerates,which even leads to serious degradation of the nano-dielectric properties.It is difficult to quantify the impact of agglomerates on materials,which is a problem that needs to be solved urgently.Capturing nanodielectric images by scanning electron microscope(SEM)is an ef-fective means to analyze agglomerations.However,due to the limitations of image pro-cessing methods,it is difficult to fully explore the value of SEM images.For example,commonly used binarization method has the disadvantages of inefficiency and inaccuracy.Motivated by the fast development of image recognition,we propose a new approach for agglomerates recognition in SEM images of nanodielectrics by semantic segmentation algorithm,which far exceeds the binarization method in terms of accuracy and general-ization ability.In this thesis,we built a SEM image dataset for the research and experiment of three semantic segmentation models.All three types of networks can recognize agglomerates in SEM images accurately and efficiently.(1)A central pixel classification network based on the pixel block is proposed in this thesis.The network crops pixel blocks with each pixel in the SEM image as the center,and classifies the pixel blocks through a convolutional neural network(CNN),thereby indirectly realizing the pixel-level classification of SEM images.The advantage of this network is that it is not limited by the small amount of SEM image data,and can accurately recognize agglomerations and matrix in the image with the MIoU of 0.837.(2)This thesis proposes a fully convolutional segmentation network based on multi-scale features.First,data augmentation methods are used to expand the number of SEMimages for meeting the training requirements of full convolutional networks(FCN).The model achieves fast end-to-end SEM image semantic segmentation for agglomeration and matrix.At the same time,pyramid pooling was used to detect multi-scale agglomeration.Finally,an image was quickly segmented by this model in 0.059 seconds,and the MIoU reaches 0.777.(3)This thesis proposes an unsupervised fitting network.This model uses superpixel segmentation methods to perform fine-grained pre-segmentation,and then uses a self-encoding CNN to fit the segmentation results of each superpixel to learn and optimize the segmentation result.This network needs 5.806s to process an SEM iamge and the MIoU reaches 0.747.According to the results of experiments,the MIoU of the binarization method is only 0.688,which is much smaller than the three networks proposed in this thesis.Further,the performance and generalization ability of three networks are evaluated.All three net-works can solve the problem of indistinguishable agglomerates in matrix with crystalline regions,and can detect agglomerates in multi-magnification SEM images.Finally,this thesis proposes a model to quantify the number and area of agglomerations based on the results of semantic segmentation.The work in this thesis has a positive effect on the fur-ther quantification of agglomerates and the development of nano-dielectrics.