Research On Image Inpainting Technology Based On Generative Adversarial Network

Image inpainting is an important research topic in computer vision.Traditional inpainting techniques focus mainly on filling the areas where some image information missed.To solve this problem,researchers have proposed many models and algorithms,which are still facing a large performance bottleneck.With the rising of deep learning,new models with higher performance than traditional models were proposed to solve the inpainting problems,but were not limited to filling blank areas.After briefly introducing algorithms of traditional image inapinting,developing history of deep learning techniques and relative application results,this thesis made detail description about the important concepts,basic principles and thought of computer vision in the field of deep learning,especially the principles of Generative Adversarial Networks(GAN).Based on relevant theoretical knowledge of deep learning,this thesis studied three different scenarios to restore damaged images,on which information partially or completely lost,into images which more in line with the vision perception system of human eyes.The main work in this thesis is as follows:(1)Restoring one-channel gray scale image into RGB image.To solve this problem,the auto-encoder model and GAN model are studied and their processing results are compared.This thesis designed a GAN model with fixed size(only height and width)of feature map in its generator to effectively alleviate two problems in existed models: color expression ability is not rich enough,the edges of different objects cannot be distinguished well,but it cannot properly deal with artifacts in low texture areas.(2)Correcting color cast in local area of images.The performance difference among Poisson image editing technique,auto-encoder model and GAN model is studied and compared.In Consideration of regular encoder-decoder model leads to fuzzy output and generator with fixed size of feature maps costs too much hardware resources,this thesis proposed to merge input image multiple times near the output layers,in this way,computation quantity was reduced and quality of restored image was improved.Experimental results showed that proposed model effectively filtered the unwanted interference from input images.(3)Restoring local areas on images where information completely lost.Based on existed images,three revision suggestions,including combining non-local attention scheme,merging input images and setting buffering layers,adding auxiliary global discriminator,were proposed to improve model performance.Experimental results showed theeffectiveness of proposed model,and the resulting image were more in line with the requirements of human vision system.