The Balanced Stacked Generative Adversarial Networks For Facial Attribute Editing

Facial attribute editing aims at manipulating single or multiple specific attributes and remaining the attribute-irrelevant image region.This technique achieves a fine-grained image manipulation,facilitating people to have a deep understanding on image data.It is widely applied into auto web glow,the generation of virtual role and data augmentation for other computer vision tasks.Although the Generative Adversarial Network(GAN)-based methods have became the dominant methods to solve facial attribute editing by the virtue of authentic edited results,they also suffer from many problems.First of all,when learning to edit multiple facial attributes,the variety of attribute combination will lead to insufficient data and produce attribute entanglement,i.e.the manipulation of originally irrelevant attributes will interfere each other.Moreover,under the condition of sample imbalance,dominant samples will dominate the study process and thus the exited methods have a poor performance on the inferior attribute values.Besides,whole image generation results in imprecise attribute editing,i.e.the attribute-irrelevant image region will be also influenced.This thesis proposes the Imbalanced Stacked Adversarial Generative Network(BSGAN)to tackle above problems.BSGAN breaks down multiple attributes editing into multiple single attributes editing task and utilizes multiple base Adversarial Generative Networks with same network structure to study independently.During the study process,samples with inferior attribute value will be assigned bigger learning weight and such samples will attach more attention.Hence BSGAN has a balanced performance on different attribute value.Moreover,only the residual image related to attribute is set as the learning target and thus the editing accuracy is improved.After study,generators of all GANs will be stacked as a stack structure,and progressively edit multiple attributes of input image.Furthermore,breaking down multiple attributes editing and residual image study ease the training of single GAN.This enables light-weight network structure and save training cost.In conclusion,the research contribution of this paper can be summarized as follows,1)Proposing BSGAN,which improves the image quality and accuracy of facial attribute editing under the condition of multiple attribute editing and sample imbalance.2)Proposed stack structure enhances the flexibility of multi-attributes editing,i.e.when considering an extra attribute,only the corresponding network not whole model needs to be trained,and then the network is stacked into the existing model.3)Experimentally demonstrating the performance of BSGAN: analyzing the image quality,editing accuracy and training cost of BSGAN comparing to popular GAN-based methods.The experiment results demonstrate that BSGAN improve the performance of facial attribute editing under the premise of achieving comparable training cost against popular methods.