Method for generating image based on multi-target sketch of progressive adversarial generation network

Abstract

The present invention provides a method for generating image based on multi-target sketch of a progressive adversarial generation network, which promotes generation of instance texture colors and promotes generation of overall image relationships by decoupling instances from the generation process of the whole image, and provides a multi-target sketch generation method based on a progressive adversarial generation network. A discriminant sketch amplification technology is used, and the sketch information of the discriminant area is acquired, so that the image generation process has more accurate structure constraint. The method provided by the invention solves the problem that the existing network neglects the distribution of instance parts in the image due to learning of the distributionof the whole image and thus generates the same texture and color on different instances. In an MS-COCO data set, the relatively high Inception Score and the relatively low Fre'chet inception distanceare obtained. The method obtains a very good result in the aspects of quality and diversity of the generated data.

Description

technical field [0001] The invention belongs to the field of computer image generation, and relates to a method for generating images based on multi-target sketches of a progressive confrontation generation network. Background technique [0002] The painting design process involves a lot of creative work. The process usually starts with sketches on paper, whereby designers and engineers share their ideas and use the sketches as a basis to create a work of art that can reproduce the real scene. Images depicting the real world often contain multiple object instances, thus image generation from sketches with multiple instances is an attractive research topic. In the past, in the field of multi-object sketch generation images, using the feature information of each sketch to perform cross-domain retrieval was the focus of research. The reference images obtained by retrieval were used to replace the instance objects in the sketches, and then a synthetic real image was obtained by...

AI Technical Summary

Problems solved by technology

However, using text descriptions to retrieve images in the network is strongly dependent on the label information of the image. If the image label in the network is inconsistent with the image, it will directly lead to an error in the final synthesis result.

Images