Uniform generative adversarial network-based multi-domain image conversion technology

Abstract

The invention discloses a uniform generative adversarial network-based multi-domain image conversion technology. The technology comprises the main contents of: training a discriminator; converting from an original domain to a target domain; converting from the target domain to the original domain; and sheltering the discriminator. The technology comprises the following processes of: learning to discriminate true and false images by using a discriminator D and classifying true images into a corresponding domain; taking the images and a target domain label as inputs of a generator G to generatea false image; under the condition of giving an original domain label, trying to reconstruct an original image by G according to the false image; carrying out continuous learning by D to discriminatetrue images and synthesized images, and carrying out continuous learning by G to shelter D; and finally trying to generate an image which is not different from the true image and can be classified into the target domain by D by G. According to the technology, a model is used for executing conversion from multi-domain images to images, so that the image conversion quality is improved, and the ability of flexibly converting input images into expected target images is provided.

Description

technical field [0001] The invention relates to the field of image conversion, in particular to a multi-domain image conversion technology based on a unified generation confrontation network. Background technique [0002] Image-to-image conversion is an effective and fast image processing and analysis technology, which converts the image defined in the original image space to another space in a certain form, and uses the unique properties of the space to more conveniently process certain images. Processing and processing, and finally converted back to the original image space to achieve the desired effect. The rapid development of electronic equipment makes the application range of image conversion more extensive. Image conversion technology can be used for the recognition and classification of human eigenfaces. First, the face image information of the target is extracted, and then the face to be recognized is projected into a new multi-dimensional face space, and the face ...

AI Technical Summary

Problems solved by technology

[0005] In view of the limitations of scalability and robustness of existing methods when dealing with more than two domains, and the problems that these methods are not efficient and effective in the task of multi-domain image conversion, the purpose of the present invention is to Provide a multi-domain image conversion technology based on a unified generation confrontation network. The process is as follows: firstly, a discriminator D is used to learn to identify real and fake images, and classify the real images into their corresponding domains; The domain label is used as the input of the generator G to generate a fake image; then, given the original domain label, G tries to reconstruct the original image from the fake image; then D continuously learns to distinguish the real image from the synthetic image, G Continuous learning to blind D; finally G tries to generate images that are indistinguishable from real images and can be classified by D as the target domain

Images