An asymmetric fine-grained infrared image generation system and method

Abstract

The invention provides an asymmetric fine-grained infrared image generation system and method, and the system comprises an encoder which is used for obtaining an infrared image, carrying out the coding of the infrared image, and obtaining the representation z of a real sample x of the infrared image; The generator is used for obtaining the representation z of a real sample, sampling the distribution P (x | z, c) to generate a generated sample x 'of the infrared image, and carrying out paired sample matching on the real sample and the generated sample; The discriminator is used for matching themean value characteristics of the real samples with the mean value characteristics of the generated samples; The classifier is used for fitting posterior probability distribution P (c | x); And the infrared image generation model is used for generating a fine-grained condition image. According to the method, the problem of image generation failure caused by unstable generative adversarial model can be solved, generation of infrared image conditions can be controlled in a fine-grained manner, and the generated infrared image is good in diversity and authenticity.

Description

technical field [0001] The invention relates to the technical field of image generation, in particular to an asymmetric fine-grained infrared image generation system and method. Background technique [0002] In recent years, image generation in the field of deep learning has always been a research hotspot, and the most outstanding representatives of generative models are Generative Adversarial Networks (GAN) and Variation Auto-Encoder (VAE). The error constructs the loss function, resulting in blurred images, while the traditional GAN ​​does not have any restrictions on how to use the representation z, which makes it difficult for model training to converge and the generated images are often unreal. Recently, many models try to improve the quality of generated samples. For example, WGAN uses EM distance as the target for training GAN, and MCGAN uses mean and covariance feature matching. They need to limit the range of parameters, so the discriminator will reduce the discri...

AI Technical Summary

Problems solved by technology

[0004] The present invention provides an asymmetric fine-grained infrared image generation system and method, the purpose of which is to solve the problems that the infrared image generation model training is unstable, the generated images are not real, and there are few types

Images