Design And Implementation Of Facial Attribute Transfer Network

Facial attribute transfer is one of the most important task in the field of computer vision,and the basic work of facial attribute research,which aims to accurately transfer target semantic attribute information in a given face image.In recent years,the development of deep learning has promoted the research of face attribute transfer.On the one hand,the current facial attribute transfer algorithms can achieve the effect of interactive entertainment by transferring several different face attributes.On the other hand,generating more face data can alleviate the problems of insufficient data and unbalanced categories to some degree in face analysis tasks.However,due to the complex face structure and fuzzy definition of face attribute,existing facial attribute transfer algorithms still have the problem of incomplete disentangling of attribute information,which leads to the change of irrelevant face information,such as face identity and artificial artifacts,when modifying the target attributes.In view of this,we study the facial attribute transfer technology from two aspects: method based on reference samples and method based on conditional vectors,respectively.The contributions are as follows:1.We propose a facial attribute transfer network based on reference examplars,Deep Semantic Disentangling Network(DSD-Net),to solve the problem of attribute disentangling.Based on the framework of conditional generative adversarial network,the DSD-Net fuses image feature and semantic attribute vectors at the channel level through low-rank bilinear pooling,to learn the corresponding relationship between them.At the same time,the DSD-Net utilize the attention mechanism to focus on the semantic information of the local region of each attribute,realizing the location and disentangling of face attributes.Furthermore,in order to improve the performance of attribute disentangling and generate high-quality face attribute images,the proposed network uses attribute classification discriminator and face reconstruction loss function for constraints.Experimental results on the Celeb A dataset and Celeb A-HQ dataset show that DSD-Net can explore the relationship among face attributes accurately transfer target face attributes.2.We propose a facial attribute transfer network taking conditional vectors as additional condition,Attribute-based Grouped Disentangling Network(AGDN),which aims to disentangle attributes and generate diverse face attribute images.The method starts from two aspects.On the one hand,we divide the dataset into groups according to labels and attribute values,which alleviates the problem of changes of irrelevant information caused by unbalanced label conditions of face attribute images.On the other hand,the AGDN adopts the cycle GAN as the basic framework and utilizes the target label-attributes as the index to map the random noise into different style codes,so as to ensure the diversity of generated images and further guide the face attribute features to realize the transfer of local attributes.Meanwhile,in order to make full use of the face features of different resolutions,we connect the encoder and decoder in space,and use the irrelevant condition discriminator to constraint training,so as to prevent the tampering of face identity and so on.Extensive experiments on the Celeb A-HQ dataset show that the AGDN can accurately transfer target face attributes and generate realistic face attribute images.With regard to two above methods,this paper further makes a comparative analysis qualitatively and quantitatively between the proposed DSD-Net and AGDN.The results show that the detail of images generated by the DSD-Net based on reference examplars are more photorealistic,while the AGDN is better in keeping irrelevant attribute information.