A Generative Adversarial Network Based Character Image Generation Method Guided by Text

Abstract

The invention discloses a character image generation method guided by text based on a generation confrontation network, which belongs to the field of computer vision. Specifically, it includes the following steps: obtain a data set of person images for training, and define the algorithm target; obtain the pose information of all images in the person image data set, and obtain the basic poses from all pose information through a clustering algorithm; The pose predictor learns from the text to the predicted pose; use the pose predictor learned in S2~S3 to predict the corresponding character pose from the text; use the character picture generator based on the generation confrontation network to generate the character picture that conforms to the text description learning, while using multi-modal errors to establish the mapping relationship between image sub-regions and text. The character image generation method guided by the text based on the generative confrontation network of the present invention has good application value in scenarios such as image generation, image editing, and pedestrian re-identification.

Description

technical field [0001] The invention belongs to the field of computer vision, in particular to a method for generating character images guided by texts based on generative confrontation networks. Background technique [0002] Text-guided person image generation is defined as the problem of simultaneously changing the pose and attributes (e.g., clothing color) of a person in a reference image to match the text description, according to the target text description. In recent years, in the fields of computer vision tasks such as specific image generation, image retrieval, and person re-identification, generative methods have played an important role in generating images with specified content, expanding data sets, and increasing the robustness of algorithms. There are two key points in this task: the first is how to predict the target pose of the character from the text, the target pose should be consistent with the text description, and serve as a guide for the transformation ...

AI Technical Summary

Problems solved by technology

Some previous methods considered the problem of changing the pose of the character, and some methods aimed at text-image generation, while fewer methods considered changing the pose and attributes of the character according to the description of the text

Images