System And Design Of CT Image Generation Of Porous Copper Foam Based On Deep Learning

Computed Tomography(CT)technology can directly obtain two-dimensional CT images of materials,while digital imaging technology can perform three-dimensional modeling of materials based on two-dimensional CT images,so as to study the characteristics of materials on the three-dimensional model.However,the small number of images obtained by CT technology and the high cost,which is not conducive to the study of the variability of materials characteristics.In order to cope with this problem,this paper uses deep neural network technology to model the porous copper foam in three dimensions,and trains the neural network with a small amount of real CT images,so that the neural network can generate a large number of similar CT images.These CT images are similar to real CT images on characteristics,but different from real CT images on structures.The flow and heat transfer characteristics of the porous copper foam are further analyzed to study how changes in the internal structural characteristics of the material change its macroscopic properties.The research work and results are as follows:1.A generative model for deep learning.Aiming at the problem of how to choose a suitable neural network model to generate a CT image of porous copper foam,we analyze the two major generative models of Deep Belief Networks(DBN)and Generative Adversarial Networks(GAN),and we built the above model by using the deep learning programming framework.2.Evaluation method for CT image generation of porous materials.Aiming at how to judge whether the CT image generated by the neural network is similar to the real CT image,we propose three considerations from the perspective of image quality,material characteristics and statistical correlation functions.By calculating the similarity between the generated CT image and the real CT image,we judge whether the generated CT image is good or bad.3.The Deep Convolutional Generative Adversarial Networks(DCGAN)based on porosity loss is proposed to generate 2D and 3D CT images of porous copper foam.Aiming at the current proposed methods for generating CT images of porous materials,there is little consideration of how to maintain the structural characteristics of the materials during the generation process,and we proposed a method for generating CT images of porous materials based on DCGAN.This method uses the DCGAN's generator to generate images and uses the discriminator to judge the true and false of the generated images.The porosity loss is added to the loss function of DCGAN,and reduces the network loss value by inputting the real CT image,thereby optimizing the generator and discrimination.Finally,the optimized generator is used to generate CT images similar to real CT images.Experiments show that the CT images generated by the improved DCGAN are similar to real CT images in terms of image quality,material characteristics and statistical correlation functions,and are better than generated CT images by using Convolutional Deep Belief Networks(CDBN).The Lattice Boltzmann method(LBM)is used to perform physical simulation on the generated CT images.Experiments show that the generated CT images can replace the real CT images to analyze the flow and heat transfer of the porous copper foam.4.A porous copper foam CT image generation system was designed and implemented.Aiming at the high computer configuration requirements of the neural network training process,the cumbersome tuning process,and the need for visualization of data,this paper designs and implements a Web system that uses the Tensor Flow deep learning framework to provide users with a neural network training interface.The Flask back-end development framework provides data analysis services,user rights management,file storage and other functions.The React front-end development framework provides services for users to interact with pages,and Mongo DB provides database support.The system provides end-to-end services for the generation of CT images of porous materials for deep learning researchers.