Face Image Inpainting Based On Generative Adversarial Networks

Facial photos,as precious identity information for humans,often encounter issues such as watermark masking,smudging,and partial information loss during dissemination,storage,and post-processing.These issues not only result in poor subjective visual effects,but also affect people’s correct analysis and understanding of real image content,adding many unnecessary troubles to our lives.Therefore,it is very important to restore damaged images to their original appearance.To solve these problems,this thesis proposes a Dense multi-scale Fusion Generative Adversarial Network(DMFB-SAM-GAN)to repair face images.This model uses a combination of dense dilated convolutions to obtain a larger and more effective receptive field,and utilizes a self-attention mechanism module to capture image global information and generates image fine detail features,and add a skip connection to the encoder and decoder of the generator to prevent mode collapse.In order to better train this efficient generator,in addition to the commonly used reconstruction loss and adversarial loss,the VGG19 feature extractor is also designed to introduce content loss and style loss,and improve the image restoration effect by adding multiple loss functions.Using a global discriminator to ensure the consistency of generated content.This thesis presents an experiment on the Celeb A image dataset,and the results show that the method can obtain pixel-level realism of face images.DMFB-SAM-GAN’s superiority in face image restoration is demonstrated by its average peak signal-to-noise ratio of 26.681 d B,structural similarity of 0.8998,and root-mean-square error of 12.293 compared with other algorithms,both qualitatively and quantitatively.The above depth generation method for face inpainting lacks proper interaction with image texture in the process of structure reconstruction,which makes it vulnerable to semantic distortion when processing damaged images.Therefore,a face image repair algorithm based on Soft-gating Dual Feature Fusion(SDFF-GAN)is proposed in this thesis.Firstly,in order to get a more reasonable result,this model adopts a texture synthesis based on structure and a structure reconstruction method based texture guidance.Secondly,to enhance global consistency,a Soft gating Dual Feature Fusion(SDFF)is designed to share and merge feature information between structures and textures,while contextual feature aggregation is utilized to generate more vivid details by modeling longterm spatial dependencies.Finally,the dual discriminator is used to repair the face image in the course of confrontation training.By conducting experiments on the Celeb A dataset,the average peak signal-to-noise ratio of the repaired image is 27.3751 d B,the structural similarity is 0.9185,and the root mean square error is 11.4503.The experimental results fully demonstrate that the SDFF-GAN network can generate higher quality repaired images.