Research On Face Image Inpainting Algorithms Based On Deep Generation Models

Image inpainting has been a hot topic in the field of computer vision.Most of the traditional image inpainting algorithms are based on image structure texture consistency.On large-area of missing face image inpainting problems,it is often impossible to obtain satisfactory repair results.With the rapid development of deep learning,the image inpainting algorithms based on the deep generation models can directly generate missing partial images,which greatly improve the effect of image inpainting.However,the deep generation models still have a lot of room for improvement.Single network structure,training process is unstable and the parameters are difficult to select.It is very challenging for the image inpainting algorithms based on the deep generation models.This thesis introduces the image inpainting algorithms based on generative adversarial networks,improves the face image inpainting method based on DCGAN,and proposes an improved face image inpainting algorithm based on LSGAN.The algorithm uses the LSGAN model as the generation network in image inpainting,which solves the two defects that low quality of the generated image and unstable training process of traditionally GAN.At the same time,the network structure and training process have been improved to improve the quality of image generated by the network.Then we use semantic loss and perceptual loss to find the best generated image which can be used to fill the damaged image,and fill the damaged image with the corresponding pixels of the generated image.Based on the public face image dataset,it can be verified that the improved algorithm has better repair ability in different types of damaged images on the subjective evaluation index and objective evaluation index.Based on the deep learning-based generative models,this thesis proposes an improved face image inpainting method combining variational autoencoder and generative adversarial network.The discriminator in generative adversarial network is integrated into the network structure of variational autoencoder.The two discriminators of the local discriminator and the global discriminator are used together for training,and the discriminant loss enhancement generator is used to generate the image.At the same time,the feature loss is introduced as the loss function of the encoder to improve the training efficiency of the generating network.When using this algorithm to repair damaged image,this image is directly inputted into the network.After encoding and decoding,the generated image is very similar to the original image.The damaged image is filled with the corresponding pixels of the generated image,and the repaired result is fine-tuned by Poisson fusion.Based on the public face image dataset,this algorithm achieves the task of inpainting large area damaged face images.Subjective and objective evaluation proves that this algorithm has strong inpainting ability for damaged images in different scenarios.