Face attribute editing method based on reinforcement learning

Abstract

The invention discloses a face attribute editing method based on reinforcement learning. The method comprises the following steps: obtaining an intermediate potential vector representation of a face in a pre-trained generator model space, and taking the intermediate potential vector representation as an initial state of reinforcement learning; inputting the obtained initial state into an Actor network, and calculating to obtain a determined action value as a direction vector of an editing attribute; adding gaussian noise gradually reduced along with the training process to the obtained action value to serve as a final action interacting with the environment; executing the adopted action for interaction to obtain a next state value; inputting the two state values into a generator model and converting the state values into a face image; utilizing a face attribute evaluator to obtain the attribute difference of the two faces and calculating an environment reward; and respectively updating parameters of the Actor network and the Critic network through gradient back propagation of the neural network by using the reward signal. According to the method, a multi-label data set required by training can be reduced, meanwhile, good attribute decoupling is provided, and the identity characteristics of an original face are kept to the maximum extent.

Description

technical field [0001] The invention relates to the technical field of computer vision, in particular to a face attribute editing method based on reinforcement learning. Background technique [0002] In reality, face attribute editing is mostly used in mobile-related face entertainment applications, such as checking one's appearance after old age, changing hairstyles, etc. In the development of deep learning, facial aging and facial makeup have become a new branch independent of face attribute editing. In addition to entertainment applications, they also have broad application prospects in case detection and face forgery. The difficulty of face attribute editing lies in two aspects: first, keep other attributes unchanged when editing a certain attribute; second, although there are many ways to edit face attributes, there are still challenges. [0003] The current mainstream face attribute editing methods are roughly divided into two types: model-based methods and additional...

AI Technical Summary

Problems solved by technology

This method is intuitively easier to understand than directly editing attributes in incomprehensible latent vectors, but creates discontinuities in attributes

Images