Research On TEM Image Segmentation Of Transmission Electron Microscope Based On Deep Learning

In the field of catalytic materials,structural information such as the shape and size of nanoparticles in supported nano-metal catalysts is a key factor affecting catalytic performance.It is related to the activity and adsorption capacity of the catalyst.How to accurately and objectively characterize the particle size information of nanoparticles is supported metal catalyst structure analysis to the primary issue of performance interpretation.Analysis based on Transmission Electron Microscope(TEM)images is the main way to obtain particle size distribution,which can provide scientific basis for researchers to grasp the catalytic performance of materials.The key step of particle size analysis is to segment and identify the image.Due to the imaging characteristics of the TEM system,the contrast between the foreground and the background of the nanoparticle image is low and the background distribution is uneven.The edges of the particles that need to be identified are blurred and the particle size is uneven.Traditional digital image processing algorithms and machine learning algorithms based on feature extraction are difficult to achieve better segmentation and recognition results in such images.With the development of artificial intelligence technology,deep learning networks are increasingly used in image segmentation and image recognition fields,such as natural scenes,medical fields,and biological fields.The main work of this dissertation first learns the main content of the dissertation by reading a large number of nanoparticle and deep learning literature,and then transfers the U-Net network suitable for medical image segmentation in deep learning to the TEM catalyst nanoparticle image,and the network is based on the addition of a lightweight module,a deep-separable convolutional U-Net network architecture that is more suitable for the hardware conditions of the transmission electron microscope is proposed.Finally,the experiment uses the core-shell structured nanomaterial as the data set,and this data set nanoparticles are small in size and complex in structure,so this dissertation mainly uses cross entropy loss function,weighted cross entropy loss function,IoU(Intersection Over Union)loss function and Dice loss for U-Net network and depth-based separable convolution U-Net network.The function is used as the optimization target to train the two networks and segment the TEM nanoparticle image.The results of the network segmentation are evaluated by the IoU and Dice coefficients and the accuracy of the segmentation.Experiments show that for the core-shell nanoparticle data set with imbalanced positive and negative samples,the deep learning network segmentation performance is better when the IoU loss function and the Dice loss function are used as the optimization targets.And use the trained network to segment the TEM image and perform statistics,which can obtain structural information such as particle size,perimeter and roundness distribution,which can extend the microscopic observation of TEM image to macroscopic structure statistics,providing feasibility for the application of deep learning in the field of catalytic materials.