Face Attribute Editing Based On Autoencoder And Generative Adversarial Network

Face attribute editing is a research hotspot in the field of image translation,which is widely used in the fields of medical treatment,beauty and criminal investigation.The purpose of face attribute editing is to change certain attributes of face image,such as age,hairstyle,expression,according to the given editing requirements,so as to get a new image with desired attributes.In recent years,the face attribute editing model has been continuously improved on the basis of the generative confrontation network(GAN),and the attribute editing ability has been greatly improved,but there are still some problems: First,the unsuitable training methods and network Settings lead to the low quality and the lack of reality in the generation of model images,and the generated face image does not meet the editing requirements.In order to solve the above problems,this thesis studies on the basis of the existing work,and the main contents are as follows:(1)In order to improve the quality of the edited image and make it more realistic,a face attribute editing network based on autoencoder and progressive generation training is proposed.Combining the autoencoder with the generative adversarial network,the encoderdecoder structure replaces the original generator,takes the real image and attribute label as input,and output the edited face image.In the process of editing using progressive training method to guide the image generation,by gradually deepening the decoder and discriminant network structure makes the model can start from low resolution to learn the characteristics of the larger scale information,with the improvement of resolution gradually shift attention to more detailed image details,avoid all features learning at the same time.The experimental results show that the network can effectively improve the generated image quality and enhance the sense of image reality.(2)In order to further improve the detail preservation ability of the model and solve the attribute entanglement problem in the editing process,the face attribute editing network based on selective transmission and convolution block attention mechanism is proposed.By adding a selective transmission layer unit(Selective Transfer Unit,STU)between the encoder and decoder,the encoder features and the decoder features are selectively connected,ensuring the attribute editing accuracy while supplementing the image details,and improving the model attribute editing ability.The convolutional block attention module(Convolutional Block Attention Module,CBAM)is added after each layer of convolution of the decoder,and the current task highlights the contribution of related features from the channel and space dimensions and suppresses the interference of irrelevant features,and realizes the decoupling of various attributes from the semantic level of the image.Experiments show that the proposed method can effectively improve the detail preservation ability of the model,and can solve the attribute entanglement problem.