Research On Image Inpainting Method Based On Improved LSGAN

Images are the fastest and most important medium for human beings to learn and acquire information besides words.Just through a picture,you can receive information such as color,person,appearance,vehicle and so on.With the successful application of deep learning technology,image-based information communication plays an increasingly prominent role in today’s society.The quality of the image is the key to determine whether the information can be accurately and comprehensively conveyed,and will directly affect the completeness of the information obtained by humans and the clarity of the information expressed by the image.Affected by factors such as weather,equipment,geographical environment,or factors such as jitter,target movement,and storage medium damage during the shooting process,the acquired images often have problems such as distortion,blurring,and damage,which affect the acquisition of image-based information and subsequent follow-up.target recognition and application.In order to solve this kind of problem,this paper combines image inpainting algorithm and GANs model,in order to solve the problems of easy disappearance of gradient,too long training time,and frequent artifacts in repair results in the image inpainting method based on GANs,the image inpainting method based on GANs is made.improvement.The main research contents are as follows:(1)An image inpainting method based on LSGAN(iLSGAN)is proposed.In the proposed algorithm,in order to effectively utilize the shallow network and deep network information,the FPN structure is used to improve the generator,and the subpixel convolution block is used to avoid the problem of artifacts in the repaired area caused by excessive artificial factors.In addition,in order to enhance the overall stability of the model,a Desne layer is used to improve the discriminator model.Experiments demonstrate the superiority and feasibility of the proposed algorithm in image inpainting tasks.(2)Through the simplification and improvement of the proposed iLSGAN model,a GAN model(i FPGAN)is proposed to improve the details of image inpainting.On the basis of the existing model,a specific residual block is designed to protect the image information while improving the generator’s performance.Repair ability,and design a multi-scale fusion network to improve the discriminator’s discriminative ability.Experiments show that i FPGAN achieves good results in both visual effects and evaluation indicators.