Image generation method based on attention mechanism and flow model

Abstract

The invention discloses an image generation method based on an attention mechanism and a flow model. The method comprises the following steps: S1, preprocessing training data; S2, enabling the preprocessed training data to enter a sub-task I, and extracting a correlation matrix by using an attention mechanism to generate an intermediate result with accurate structure information; S3, enabling the training data to enter a sub-task II, and predicting a flow field by using a flow model to generate an intermediate result with rich detail information; S4, inputting the correlation matrix generated by the sub-task I and the flow field generated by the sub-task II into a sub-task III in a one-to-one correspondence manner under different scales, and generating a weight map by using a content awareness network; and S5, generating results of the subtask 1, the subtask 2 and the subtask 3 and training data enter a main task, a human body image is generated by using an encoder-decoder network, and the image is output. Structural information and detail information can be considered, so that better generation quality is obtained.

Description

technical field [0001] The invention relates to the fields of image generation and generation confrontation network, in particular, an image generation method based on attention mechanism and flow model. Background technique [0002] With the continuous development of deep learning and neural networks, the fields of computer vision and image processing have received extensive attention. Among them, the pose-guided human body image generation task that has emerged in recent years is a challenging and practical topic in the field of computer vision. Its core task is to transform the characters in a given image into the appearance in the target pose through a series of spatial transformations. This task has many difficulties, such as incomplete structural information in the generated image, lack of texture information, and so on. [0003] When the human body image generation task was first proposed, most methods used a simple encoder-decoder structure. However, due to the la...

AI Technical Summary

Problems solved by technology

This method can usually generate a more accurate structure, but it is difficult to preserve the clear texture of each semantic part of the original image

[0006] All in all, although the current methods can perform better in some aspects, most of them are difficult to take into account both texture information and structural information.

Images