Research And Implementation Of Image Generation Algorithm Based On Optimal Transport

With the development of computing devices,deep learning has gained widespread attention in the near future,and its efficiency and high accuracy have surpassed the traditional algorithms to a certain extent.However,the inexplicability of deep learning makes its application in industries with high reliability requirements,such as medical,military,and financial,severely restricted.At the same time,another constraint that hinders the widespread use of deep learning is the lack of training data.In order to propose a reliable and interpretable deep learning image generation model,and at the same time in order to avoid the problem of mode collapse and mode confusion in the generation model,this work applies the optimal transport theory to the image generation model.We use the Monte Carlo method to extend the semi-continuous optimal transmission algorithm to any dimension,making it possible to combine it with deep learning models.We use the deep learning tensorflow-gpu framework to build and implement multi-GPU sampling and parallel computing,which increases the number of Monte Carlo sampling points,which improves the convergence accuracy of the optimal transmission algorithm and also improves Computing efficiency.In the generation model,we propose a partially interpretable deep learning model.In this paper,we apply the optimal transport algorithm in the field of deep learning image generation,and propose a semi-transparent generation model,Optimal transport-Autoencoder generation model and Optimal transport-hidden space interpolation generation model,to make it in applications in areas with high reliability requirements are possible.At the same time,we tested and analyzed the image data generated by the two models.The optimal transmissionhidden space interpolation generation model solved the problem of mode collapse and mode confusion in image generation,which was verified by analysis of the model test results.