Research On Image Generation Algorithm Based On Variational Auto-Encoder

As one of the most widely used algorithms in unsupervised learning,generative models can effectively capture the deep structure of data and be used to generate new data.It has important applications in image reconstruction and image denoising.Variational autoencoder,as one of the most important generative models,has the advantages of easily importing statistical priors,but its image generative ability needs to be further improved due to the prior and structure limitations.Therefore,the variational autoencoder with optimizing Gaussian mixture prior and the implicit-prior variational autoencoder based on attention mechanism are proposed to improve the image generative ability.The latent variable prior of the variational autoencoder often uses a simple standard normal distribution due to the convenience of calculation,but there is a problem of underfitting.The Gaussian mixture model can be used to construct more complex prior distribution,but there is a problem that the Kullback-Leibler distance between the two mixture Gaussian probability density functions has no closed-form solution.To solve this problem,a variational autoencoder with optimizing Gaussian mixture prior is proposed.This method first uses the greedy algorithm to solve the approximate solution of the Kullback-Leibler distance to define the approximate solution of the variational lower bound of the loss function,and then uses the Kullback-Leibler distance between the posterior distribution and the prior distribution to realize the data-based prior Iterative optimization of the distribution.The experimental results on the datasets MNIST and Omniglot show that variational autoencoder with optimizing Gaussian mixture prior can effectively improve the lower bound of the model's log-likelihood and improve the model's generative ability.Considering that the features extracted by the neural network model of the variational autoencoder are the overall features of the data set,but for a single sample,there are not only effective features,but also invalid features,and it is these invalid features that greatly limit the generative ability of the model.To solve this problem,the implicit prior variational autoencoder based on attention mechanism is proposed.This method parallels a new branch which is the same as the original structure,adds a sigmoid activation function after the new branch,and then multiplies it pixel by pixel with the original structure,so that the network can achieve a dynamic feature selection function in each spatial location and each channel,retain the interpretive features and remove the redundant features.At the same time,the Kullback Leibler distance is rewritten to convert it into low dimension,and the Kullback Leibler distance is estimated implicitly by binary classifier.The experimental results on MNIST,omniglot,Frey faces and onehot data sets show that the implicit prior variational autoencoder based on attention mechanism can improve the lower bound of log likelihood of the model and effectively improve the generative ability of the model.