Research On Image Generation Algorithm Based On Autoencoder

In recent years,images have gradually become the main carrier of information dissemination,and cross-domain image generation has gradually attracted people's attention.The development of deep learning has provided technical support for cross-domain image generation.However,early work often relied on paired data sets,and it is difficult to find one-to-one paired images in real life,so this limits the generalization ability of cross-domain image generation tasks to a certain extent and is difficult to meet The needs of real life.In addition,most of these cross-domain image generation work can only achieve one-to-one image generation,which cannot meet the one-to-many generation requirements.This article focuses on the above problems,and is committed to building an unsupervised(no paired data)one-to-many image generation algorithm.In view of the above problems,this paper proposes a cross-domain image generation algorithm based on self-encoding.Assuming that the cross-domain image has independent style attributes and consistent content attributes,first use the encoder to encode the cross-domain image to obtain its content attributes and style attributes.For the independent style attributes,the variational autoencoder is used to prepare To make it possible to meet the Gaussian distribution set as far as possible;for consistent content attributes,the adversarial self-encoder is used to perform adversarial learning on its domain labels and category labels to fit the prior distribution of content attributes.Finally,the fitted style attributes and content attributes are randomly sampled,and then stitched to achieve cross-domain image generation.The algorithm of this paper has conducted supervised and unsupervised experiments on the four data sets of MNIST,SVHN,VIS-NIR and Edges-Shoes,which fully verified the effectiveness of this algorithm.