Research On Image Inpainting Method Based On Deep Learning

Digital image inpainting refers to predict the missing area of damaged image via the neighborhood information according to certain rules,so that the observer cannot find obvious repair traces.Due to the limitation of the repair strategy and the ability of model representation,traditional image inpainting algorithms cannot get satisfactory repair results in some complex situations.Deep learning,with its powerful ability of feature learning and extraction,has made great achievements in image inpainting,especially in semantic image inpainting.However,the existing deep learning-based image inpainting algorithms still have some deficits to be improved,such as the instability of training and insufficient of accuracy.This paper is devoted to the study of image inpainting using deep learning.Aiming at the defects and problems of the existing image inpainting methods,two algorithms based on deep learning are proposed.The specific research contents are as follows:(1)Aiming at the defect of the inpainting strategy of traditional image inpainting algorithms and the deficiency of model representation ability,an image inpainting algorithm based on deep Auto-Encoder is proposed.In this algorithm,a multi-task deep Auto-Encoder is trained,which can simultaneously extract image features and predict missing pixels.Consider that the natural images have local correlation and non-local self-similarity,we use the local information to predict the missing pixels and the non-local information to correct the predictions,which can improve the accuracy of the predictions.In addition,in order to alleviate the error propagation problem,we proposed a novel adaptive inpainting sequence based on the joint credibility of support area and reconstruction.(2)In view of the defects of the existing image inpainting algorithm for semantic information loss,a progressive image inpainting algorithm based on Generative Adversarial Networks(GAN)is proposed.In this algorithm,the repair of the entire damaged region is divided into several sub-stages.At each stage,only a part of the damaged region is repaired,and the Long Short-Term Memory network(LSTM)is used to connect these sub-stages.In this way,the training process of the network is more stable and the image inpainting is more precise.The neural network used in each sub-stage consists of two parts: the generative part and the discriminative part.The generative part is implemented with an encoder-decoder neural network,and we concatenate somecorresponding layers of the encoder network and decoder network through “shortcut”,which can alleviate the information lost in the process of down sampling,enhance the generation ability of the network and reduce the risk of gradient vanishing.In addition,this network uses two discriminative networks: global discriminative network and local discriminative network,which can improve the repair effect of the damaged area.Experimental results show that compared with the comparable algorithms,the two algorithms proposed in this paper get better repair results in terms of objective and perceptual metrics.