Image generation method and device

Abstract

The invention provides an image generation method. The image generation method comprises the steps of obtaining a to-be-generated signal; inputting the signal to be generated into the first generator model to obtain a generated image output by the first generator model, wherein the first generator model is obtained by performing adversarial training based on a compression adversarial network formed by the first discriminator model and the pre-training discriminator model, the pre-training discriminator model is a pre-trained discriminator model, the first discriminator model and the pre-training discriminator model are respectively used for discriminating the output of the first generator model to obtain a discrimination signal, and the discrimination signal acts on the adversarial training process of the first generator model. According to the invention, starting from a teacher-student learning normal form, supervision of the pre-training discriminator model is introduced into the adversarial network, so the performance of the lightweight generative adversarial network is improved, and the lightweight generator has similar image generation capability to an original generator.

Description

technical field [0001] The present disclosure relates to the technical field of image generation, and in particular to an image generation method and device. Background technique [0002] Generative adversarial networks have important applications in image generation. The parameters of existing generative models for high-resolution images are often very large, which brings great challenges to the application of generative models in actual production environments. The current common practice is to compress the network structure of the generative confrontation network to reduce the model parameters and improve the usability of the model in the actual production environment. [0003] At present, the adversarial network with compressed structure has the disadvantages of high computational complexity of the compressed generator and poor performance of the compressed generator. Most of the existing GAN search methods use teacher generators to provide supervision signals, or buil...

AI Technical Summary

Problems solved by technology

[0004] The present disclosure provides an image generation method, which is used to solve the defect that the generation model of high-resolution images has a huge amount of parameters in the prior art, and the quality of the image generated by the compressed generation model is poor, and realizes the application of lightweight compression generation confrontation network The purpose of generating high-quality images

Images