Research On Progressive Image Generation Model Based On Attention Mechanism

In recent years,with the rapid development of computer technology and the neural network,as a more complex learning component of machine learning,unsupervised learning has gradually become the research hotspot of machine learning as the research fruits of supervised learning have mostly been picked off.As one of the techniques to try to solve the problem of unsupervised learning,generative model has been developed rapidly.The main purpose of generative model is to generate more samples which is similar to the data of the specified database,but the generated samples are required to be diverse at the same time,and cannot match the date in the specified database.With generative model,we can generate some new image data which never been seen from an image database,and a massive database can be generated from the limited size of the date.It increases the sample size of data,and provides more abundant training data for machine learning.Among them,deep neural network models,such as generative adversarial network(GAN)and variational autoencoders(VAE),have achieved better results in image generation.This paper mainly studies progressive image generation method based on attention mechanism,combines neural network with the human visual attention mechanism,based on the sample data of codec,reconstructing similar but different new image with the original sample.The main research and work of this paper are as follows:1.This paper constructs an image generation model based on GAN.The model consists of two parts:the generation model and the discriminant model.It is a learning framework,which is actually a simulation game between the generation model and the discriminant model.The purpose of the generation model is to try to imitate,model and learn the distribution of real data,and the discriminant model is to distinguish one input data from the real data distribution or generated from a model.Through the constant competition between these two internal models,the generation ability and discrimination ability of the two models can be improved.2.This paper builds an image generation model based on VAE.The model encodes the input images through the encoder,and then through a decoder to output generated images.It depends on the traditional probability theory model framework,through some proper joint distribution probability approximation,simplify the whole learning process,which makes the model to explain the observed data very well.3.We proposes an image generation model based on the attention mechanism and the recurrent neural network.The model on the basis of VAE joins the attention mechanism and RNN.When computer reconstructs images,it is able to draw as much emphasis as human painting,rather than like a printer.The experiment of image generation is conducted on MNIST and CIFIA-10 datasets in this paper.The result verifies the effectiveness of progressive image generation model based on attention mechanism.Through comparing the three generative models,our model has the best performance and the generative image quality is the highest.