Decoder training method, high-resolution face generation method, device and computer equipment

Abstract

The invention provides a decoder training method which comprises the steps of acquiring a first face image which is a low-resolution face image with resolution that is lower than a first preset value;coding the first face image by a decoder, and forming a high-dimension vector which corresponds with a first face image characteristic, and the dimension of the high-dimension vector is higher than asecond preset value; generating a second face image by a generator according to the high-dimension vector, and the second face image is a high-resolution face image with resolution which is higher than a third preset value; determining the value of a loss function according to the second face image, and adjusting the decoder parameter according to the value of the loss function; and when the value of the loss function satisfies a preset condition, obtaining the trained decoder. According to the method, after the high-dimension vector which is output by the decoder according to the obtained low-resolution face image is input into the generator, the high-resolution face image which approaches the low-resolution face image can be generated.

Description

technical field [0001] The invention relates to the field of image processing, in particular to a decoder training method, a high-definition human face image generation method, device and computer equipment. Background technique [0002] With the development and application of artificial intelligence technology, it is necessary to generate high-definition face images in many scenes. In the prior art, there is a method of using random variables to generate high-definition face images. For example, the photo-realistic face generation software released by Nvidia uses the Generative Adversarial Nets (GAN) method, using two neural networks to train each other. The image generation network tries to generate synthetic images that are indistinguishable from real photos, while The adversarial network tries to tell the difference, so that after a few weeks of training, the image-generating network can generate images of faces that are recognizable as real ones. [0003] However, th...

AI Technical Summary

Problems solved by technology

[0003] However, the existing technology uses random variables to generate high-definition face images, and the generated faces are random, but in practical applications, it is often necessary to generate high-definition pictures of specified faces. Face image methods cannot meet this need

In the existing technology, there are two latest technologies to process blurred face images into clearer face images. One technology uses image super-resolution reconstruction method, and the other technology uses conditioning network (Conditioning Network) and prioritization The network (prior network) priority network will make a decision between low-resolution and high-resolution photos, and fill the high-resolution photos according to the probability. The existing technology is to fill the low-resolution images to supplement the low-resolution images. The new details that the image does not contain, so the details added by the existing technology in the process of converting low-definition pictures to high-definition are just a "guess"

Images