Research On Image Completion Method Based On Generative Adversarial Networks

The research of image completion model is helpful to help people restore or modify images more conveniently.At present,the researches on image completion methods are most commonly based on deep learning models,and the results are more excellent.However,due to the variety of missing image types,most image completion models cannot effectively fill and complete image with boundary missing.At the same time,image completion models based on deep learning require a large amount of training data and largescale training as support,which increases the limits of the model in practical applications.Therefore,how to construct an image completion model that can effectively complete a variety of missing situations without the need for large-scale training is a hot issue worthy of research.In view of the above problems,this thesis proposes an image completion model based on generative adversarial networks.By adding auxiliary image processing algorithms,the model can quickly converge during training;then the structure of the mode is optimized to improve the operation efficiency and completion results.The specific research content is as follows:1.According to the importance of known pixels,this thesis designs a pre-processing algorithm based on region growth,which highlights pixels that are similar to and connected to known pixels at the boundary of the missing region,to help the model understand the image information and accelerate the convergence speed.According to the comparative experiment,our model is more effective for boundary missing.2.In order to further improve the efficiency of the model,Attention mechanism module is used to improve the speed of algorithm.Attention mechanism module sets attention weight to each layer of neural network in the encoder,which is part of the condition generator.According to different network parameters,attention weight also changes accordingly.In this way,the model can clearly focus on the important area at each layer of feature extraction,and avoid missing pixels from interfering with results.3.As image sharpness is a kind of high-level semantics that is difficult to quantify,this thesis designs an acuteness discriminator based on the cloud model to identify the sharpness of the image.The cloud model is used to transform the uncertain concept of image sharpness into a quantitative problem,which constrains the sharpness of the generated results.Experiments show that our completion results generated are obviously better than other methods in detail processing.