Image Generation Based On The Gated Self-Attention Auto-Encoder

Variational Auto-Encoder(VAE)is one of the generative models and has been successfully applied in the field of natural language processing,such as text generation,machine translation and text classification.However,Variational Auto-Encoder has the problem of prior collapse,and when the Variational Auto-Encoder is applied to image generation tasks,it tends to generate images with blurred borders.In order to solve these two problems,this paper analyzes the causes of the problems in detail,and proposes an abstract framework for Auto-Encoder.And based on the abstract framework of Auto-Encoder,a new generative model—Self-Attention Auto-Encoder(SelfAttention AE)is proposed.The shallow layer of the neural network can learn simple and specific features,such as border features and color features.As the number of network layers deepens,the learned features become more abstract.Self-Attention AE uses mutual information to maximize the similarity between the data feature generated by the encoder and the shallow feature thereby preserving the edge features of the images,and uses self-attention to weight the features of different dimensions to increase the correlation of similar features.Experiments show that the Self-Attention AE can generate higher quality samples than VAE and Wasserstein Auto-Encoder(WAE),as measured by the Freshet Inception Distance(FID)score and Sharpness score.To further improve the quality of the images generated by the model,this paper improves the convolutional layer of the encoder based on the Self-Attention AutoEncoder,and introduces the gated attention mechanism into the model.Experiments show that when the gated attention mechanism is added,the quality of the images generated by the model will be further improved.Although the Self-Attention AE and the Gated Self-Attention AE can improve the quality of image generation,there is still a gap compared with Generative Adversarial Network.In future work,I intend to continue to explore the Auto-Encoder framework and improve the VAE so that it can generate higher quality images.At the same time,I plan to apply the Self-Attention AE and the abstract framework for Auto-Encoder to the field of natural language processing to explore whether they can play a role in natural language processing tasks.