Generative Adversarial Networks Based Face Atrribute Translation

Image translation task is an important research topic in generative adversarial network.In real life,image translation has a lot of practical application needs.In this paper,we focus on the task of face attribute translation in image translation task.Because human vision is sensitive to the irregularity and deformation of human faces,it is not easy to transform the face attributes,especially how to edit the generated face image freely and flexibly according to the user’s will.The original end-to-end training model are often limited to large amounts of labeled datasets,which is unable to edit face freely and flexibly.The latent space of mage generation model often contains rich semantic information.The task of face transformation on the operation of latent space has gradually become a research emphasis.This paper is mainly to explore the method of region editing on face based on the latent space of generative adversarial network.Region face editing task is a face attribute translation task that only focuses on the region of interest in the face.The model used in this paper introduces the encoder for encoding a specific picture.The method of this paper is mainly aimed at doing research on the latent space based on this model and studies various possibilities of region editing.We realize region editing task by two different ideas of region feature collage and region vector semantic editing.Region feature collage uses target face image fused with the specific region of the reference image to achieve the attribute translation from the specific region of the reference image.A new normalization method is introduced to make the latent code of reference image region adaptive to the latent code distribution of target image region.The effectiveness of our new normalization method has been proved by ablation experiments.At the same time,we have carried out the experiments of facial features collage and face expression translation and achieved excellent results.Region vector semantic editing is to find the disentangled semantic operation vector and realize that when "walking" on the semantic vector,the specific semantics of the region of the face image can be changed.In this paper,we propose a new method called patch PCA to calculate the semantic operation vector,use unsupervised and weakly supervised methods to disentangle the latent space and get specific semantic manipulation vectors.The weak supervised patch PCA method in the region of latent code can obtain the specific semantic manipulation vector directionally and perform more flexible attribute operation on the local area.We carried out a series of experiments in mouth region,eye region and eyebrow region for disentanglement and achieve a specific vector to manipulate the eyebrow.Excellent experimental results demonstrate the effectiveness of our method.