Unsupervised image-to-image translation method based on content style separation

Abstract

The embodiment of the invention discloses an unsupervised image-to-image translation method. A specific embodiment of the method comprises the following steps: acquiring an initial image, and zooming the initial image to a specific size; performing spatial feature extraction on the initial image through an encoder to obtain feature information; inputting the feature information into a content style separation module to obtain content feature information and style feature information; generating reference style feature information of the reference image in response to the acquired reference image, and setting the reference style feature information as Gaussian noise consistent with the style feature information in shape in response to the fact that the reference image is not acquired; inputting the content feature information and the reference style feature information into a generator to obtain a target image translating the initial image into a reference image style; and zooming the target image to a size matched with the initial image to obtain a final target image. The implementation mode can be applied to various different advanced vision tasks, and the expandability of the whole system is improved.

Description

technical field [0001] Embodiments of the present disclosure relate to the technical fields of computer vision and image processing, and in particular to an unsupervised image-to-image translation method. Background technique [0002] Image-to-image translation has attracted much attention since it can learn the mapping between different visual domains. In the current social media or chat software, many chatting objects are converted into cute animals, and the converted animal expressions are consistent with the original object during the chatting process. Or change the style of the chat background while maintaining the spatial structure information of the background. Or it is necessary to exchange the virtual scene with the real scene in the virtual-real scene of the augmented reality application, etc., which are typical applications of image-to-image. Pix2pix (pixels to pixels, pixel-to-pixel conversion tool) is the first to use conditional generative adversarial network...

AI Technical Summary

Problems solved by technology

Therefore, the disadvantages of this type of method are: 1. It cannot guarantee that the features of the content can effectively focus on meaningful objects in the image

2. Style features often focus on the appearance of the entire image, and cannot effectively extract different styles of different objects

The disadvantage of this type of method is that different network architectures need to be designed for specific high-level visual tasks. If the corresponding high-level visual information cannot be provided on the new data, it is difficult to extend the same method to multiple image-to-image translation tasks.

Images