Research On Deep Generative Model Based Image Inpainting Algorithm

Image inpainting is an important part of research in the field of computer vision,and its purpose is to restore the defective part of the image by using the existing information in the image.With the development of deep learning,image inpainting technology based on convolutional neural networks has important applications in the fields of criminal detection,cultural relic protection,and special effects of film and television.However,the existing image inpainting methods have some disadvantages such as insufficiently clear image inpainting results and insufficient results diversification.Based on this,this paper takes the face image as the research object,and proposes a new face image inpainting method based on variational autoencoder(VAE)to obtain a clearer and more diverse repaired image.This paper firstly introduces the research history and current status of traditional image inpainting methods and deep learning-based image inpainting methods,focuses on deep learning-based image inpainting methods,and introduces the research status of three mainstream deep learning generation models;then introduces the theoretical basis of VAE,Generative Adversarial Networks(GAN)and Poisson image editing involved in this paper,and proposes an improved image inpainting model based on existing methods.Aiming at the problem that the existing method of repairing the face image is not clear,this paper uses VAE to repair the image,and introduces the design of the GAN discriminant network and loss function to improve the clarity of the repaired image.In order to obtain a better inpainting effect,dynamic programming is used to find the minimum error segmentation boundary between the generated image and the image to be repaired,and Poisson image editing is used to obtain a seamless fusion result.Finally,this paper constrains the latent variables in the image inpainting model so that the dimensions are as close to each other as possible.Independently,implement feature disentanglement operations to obtain images with specific attributes and increase the diversity of repaired images.This paper performs experiments on the CelebA dataset.Compared with the classical method,the method in this paper has obtained good image inpainting results in both qualitative and quantitative comparisons.And the effect of image fusion on image inpainting is analyzed through experimental comparison.At the same time,by explicitly controlling different dimensions of the latent variables,the face image inpainting results with different attributes are displayed.