Seasonal style conversion model and method for image named MSGAN

Abstract

The invention relates to a seasonal style conversion model and method for an image named MSGAN. The seasonal style conversion model comprises a generator G, a generator F, a first true and false discriminator DG, a second true and false discriminator DF and a seasonal discriminator DS. The generator aims to convert an input image into other specific seasonal styles. The true and false discriminator is used for distinguishing whether the image is a composite image. And the season discriminator performs season classification on each synthesized image and each real image. The two discriminators can provide guidance for the generator respectively. In order to give a correct optimization direction of the network, the MSGAN respectively uses style loss, structural similarity loss and color lossto improve the generation capability of the generator. Moreover, the MSGAN uses the saliency information of the image for the first time to guide the image style conversion task, so that the image style conversion result better conforms to the real situation of human eyes.

Description

technical field [0001] The invention relates to the technical field of image processing, in particular to an image season style conversion model and method called MSGAN. Background technique [0002] In recent years, cartoons, animations, and 3D movies have become popular art forms. Their advantage lies in that they can be used to design characters and scenes through computer software and use them to perform scripts. In animation works, designers not only need to design the characters, but also need to draw the scenes that appear in the works. Due to the complexity and diversity of the storyline, the season style of the same scene will often change seasons with the plot. Not only animation works, many regular movies or TV series also need such special effects. However, it is time-consuming for designers to manually use computer software to complete the seasonal switching task and requires a lot of related software skills. Therefore, it is an important task to design a spe...

AI Technical Summary

Problems solved by technology

[0004] Although generative adversarial networks have achieved significant success, current methods do not perform well in the task of seasonal style transfer.

There are three main reasons for this: First, season conversion is different from other style conversion tasks. Sometimes it is necessary to add or reduce certain elements in the original image. For example, images under winter scenes need to add some snow and reduce some leaves, which is a difficult point in the style transfer task

Second, generally speaking, the main feature of seasonal changes is the transformation of colors, and most image style transfers focus on adding image texture to the original image, while ignoring color changes.

Third, when performing seasonal style conversion, different content is affected to different degrees by seasonal changes. For example, leaves appear green and yellow in spring and autumn respectively, while the color of tree trunks does not change much, but the traditional image style migration It is difficult for the algorithm to identify the content of the image

Images