Self-supervised attribute controllable image generation method based on deep twin network

Abstract

The invention discloses a self-supervised attribute-controllable image generation method based on a deep twin network, and belongs to the field of computer vision image generation. The method is based on a twin network idea, takes cosine similarity as a distance measurement method, restrains the similarity between two enhanced samples of the same image, and further stabilizes a network training process by using a gradient stopping strategy, thereby providing differentiated representation of category attributes in real data for a twin encoder network, and improving the accuracy of the twin encoder network. And the learned category representation is used for an image generation process, a generator is induced to realize category attribute control, meanwhile, common mean square error loss in other methods is deleted, and the model optimization difficulty is reduced. According to the self-supervised attribute control scheme based on the deep twin network, the class attribute control generation level of the generative adversarial network can be remarkably improved, the image generation quality of an existing method is improved, and a more excellent effect is achieved through fewer constraints and shorter training time.

Description

technical field [0001] The invention belongs to the field of computer vision, mainly relates to the task of controllable generation of category attributes of images, and is mainly used in cultural entertainment industry, industrial image data expansion, machine vision understanding and the like. Background technique [0002] Image generation technology is a technique that solves the maximum likelihood estimation of image data distribution by establishing a mathematical model, and samples from the estimated distribution to generate new images that are similar to the original data but do not exist in the original data. In recent years, with the rapid development of deep learning, a number of image generation technologies based on deep neural networks have emerged, and have achieved image generation effects comparable to real images. According to whether to directly calculate the maximum likelihood estimation of the real image distribution, it can be divided into two categories...

AI Technical Summary

Problems solved by technology

[0009] The invention is a self-supervised attribute-controllable image generation method based on twin networks, which mainly solves the problem of insufficient category attribute control effect

Images