An image local information migration method based on GAN and Self-Attention

Abstract

The invention relates to an image local information migration method based on GAN and Self-Attention, which comprises the following steps: 1) acquiring an image migration task; 2) according to that image migration task, a trained neural network is used for image local information migration, and the migrate image is outputted; The neural network comprises an image generator and a discriminator, wherein the image generator comprises a convolution network and a deconvolution network, wherein the convolution network convolves pictures into hidden feature spaces, andthe Self-Attension model is usedto exchange part of hidden features, and that exchanged features are de-convoluted into new pictures through the de-convolution network so as to realize the feature transfer of local information of an image; The discriminator is used for discriminating the true-false relationship of the generated picture. Compared with the prior art, the invention has the advantages of avoiding the influence of the local migration and the global pixel of the image, solving the interference factors of the local migration of the image to the whole image, etc.

Description

technical field [0001] The present invention relates to the field of deep learning, relates to an image information processing method, in particular to a method for image local information migration based on GAN and Self-Attention. Background technique [0002] With the continuous development and evolution of machine learning methods, big data has become a treasure trove of knowledge in various industries such as academia and industry. However, at present, there are unbalanced data categories or even missing data in many fields, which seriously affects the stability and robustness of machine model training. How to generate high-quality missing data from existing datasets has become the primary problem to be solved. [0003] At present, applying Generative Adversarial Network (GAN) to generate new image data has become a research hotspot in the field of deep learning. Among them, applying image-2-image translation technology to study new image generation is usually an effec...

AI Technical Summary

Problems solved by technology

However, ELEGANT did not control the local information of the image and the interference factors of the pixels of the whole image.

Therefore, it is easy to affect the generation of other pixel information in the image field during the migration process

Images