Image Restoration Algorithm Based On U-Net Networks And It Application On Yi Character Restoration

Image restoration refers to filling the missing area with reasonable content,so that the restored image can obtain reasonable overall structure and fine texture details.Whether it is based on traditional image restoration methods or deep learning based image restoration algorithms,when repairing large areas of missing areas,the rationality of the global semantics,the integrity of the overall structure,and the fineness of local texture details cannot be obtained satisfactory repair results.At present,there are very few studies on the application of deep learning-based image restoration to handwritten Yi character restoration,and the application of computer deep learning to handwritten Yi character restoration will provide a technical platform for later studies and work related to Yi character restoration.In order to carry out reasonable structural restoration and fine texture filling of large-area missing areas in the image,this paper focuses on some key image restoration techniques,including the inpainting algorithm for depth feature rearrangement based on double shift network and Two-Stage network image inpainting algorithm based on BDCN（Bi-Directional Cascade Network for Perceptual Edge Detection）and U-Net incomplete edge generation,and applied to repair script Yi character.The details are as follows:1.The inpainting algorithm for depth feature rearrangement based on double shift network is proposed.This method uses double shift network down-sampling to extract features from the input image firstly,and then combines the input information of up-sampling and information of each layer of down-sampling to restore image texture details.After that,we introduce two special shift-connection layer to the U-Net architecture,the encoder feature of the known region is shifted to serve as an estimation of the missing parts.A content loss is introduced on decoder feature to minimize the distance between the decoder feature after fully connected layer and the ground-truth encoder feature of the missing parts,to ensure the accuracy of the semantics of the synthesized image and the rationality of the structure.On the Places datasets,the proposed algorithm is compared with the existing classic algorithm.The subjective and objective experimental results demonstrate the proposed algorithm can obtain clear,fine-detailed and visually reasonable results in the restoration of large missing areas images,which is superior to contrast algorithm.2.Two-Stage network image inpainting algorithm based on BDCN（Bi-Directional Cascade Network for Perceptual Edge Detection）and U-Net incomplete edge generation is proposed.The algorithm first extracts edges based on the BDCN network,each layer of the network obtains a specific-scale edge feature through learning,and subsequently combines these multi-scale features to generate a complete and reasonable edge.In the first stage,based on the edge generation network of the U-Net network architecture,down-sampling is used to extract the features of the missing image edges.Thereafter the information inputted by the up-sampling layer and the information of each layer of the down-sampling are combined to restore the image edge texture details.In the second stage,image restoration network combines the incomplete image and the predicted complete edge by using the hole convolution for down-sampling and up-sampling.Finally the residual network reconstructs the missing image with rich details.The proposed algorithm is compared with the existing classic algorithm on the Places datasets.The subjective and objective experimental results demonstrate that the proposed algorithm can obtain reasonable results and fine texture details in the restoration of large and irregular missing areas in images,and it performance is superior to those of the contrast algorithms.3.Combined self-made and public datasets to create handwritten and printed Yi character datasets,use small masks and large masks to simulate the degree of damage to Yi character,The inpainting algorithm for depth feature rearrangement based on double shift network and Two-Stage network image inpainting algorithm based on BDCN and U-Net incomplete edge generation are applied to the handwritten and printed Yi character restoration.The proposed algorithm is compared with the existing classic algorithm on the existing datasets.The subjective and objective experimental results demonstrate that: 1.The two proposed image restoration algorithms have good results in achieving a certain degree of Yi character restoration.2.The proposed algorithm is better than the comparison algorithm in the restoration of handwritten and printed Yi characters under different masks.