Hair Attribute Transfer via Deep Feature Fusion

To tackle the problem that existing attribute transfer methods can’t transfer hair attributes effectively,a method of hair attribute transfer based on deep feature fusion is presented.This method includes three subnetworks which are responsible for feature extraction,attribute vector extraction and image synthesis.Firstly,feature extraction network extracts features from original images,and keeps the identity of original images unchanged by adding a reconstruction loss.At the same time,attribute vector extraction network constructs the mapping model of hair features and hair attributes,and generates the attribute vector.Finally,the synthesis network takes the fusion result of image features and the attribute vector as input,and generates final results.Various attribute transfer experiments on FFHQ show that the proposed method can effectively transfer hair attributes and generate high-resolution results.Experiments on Celeba show that the proposed method can achieve better visual quality than existing popular attribute transfer methods.