Face image diversified restoration method based on sample guidance

Abstract

The invention provides a face image diversified restoration method based on sample guidance. A mapping network, a style network, a generator network and a discriminator network are included. In each iterative training, the mapping network maps random Gaussian distribution to a random style; the style network encodes the style of the sample picture to obtain a sample style; the generator network performs global style extraction on an input image, and then embeds a random style, a sample style and a global style into a decoder in a generator to generate a face repair result containing sample attributes. And calculating a loss value by combining generative adversarial loss, spatial variant perception loss, identity loss and attribute consistency loss, performing back propagation, and adjusting parameters of the mapping network, the generator network and the discriminator network. Repeating the steps until the training is finished, and selecting an optimal network parameter as a model generation parameter; according to the invention, a high-quality sample-guided face image diversified restoration method can be realized.

Description

technical field [0001] The invention relates to the technical field of image restoration, in particular to a method for diversifying face image restoration based on sample guidance. Background technique [0002] In image inpainting methods, traditional inpainting methods can use similar patches around the image to fill in the missing image regions, but these methods cannot capture the semantic relationship of the image well, which leads to the poor performance of these methods in complex images. [0003] Although the existing deep learning-based inpainting methods can generate satisfactory results in the missing areas, these methods have the following problems: First, the image inpainting results of most methods tend to be single, and the user cannot choose the desired image. Repair results; second, for methods that can output multiple repair results, although the results are diverse, the user cannot guide the desired output effect; third, some repair methods can introduce s...

AI Technical Summary

Problems solved by technology

[0002] In image inpainting methods, traditional inpainting methods can use similar patches around the image to fill in missing image regions, but these methods cannot capture the semantic relationship of images well, resulting in these methods often perform poorly in complex images

[0003] Although the existing inpainting methods based on deep learning can generate satisfactory results in the missing area, these methods have the following problems: First, the image inpainting results of most methods tend to be single, and users cannot choose the one they want. Repair results; second, for methods that can output multiple repair results, although the results are diverse, users cannot guide the desired output effect; third, some repair methods can introduce some auxiliary information, such as sketches, strokes and feature keys points to control the face generation

However, due to these methods relying too much on prior information, the repair results fit the prior and limit the diversity of repair.

Images