Research On Image Inpainting Technique Based On Edge Prior And Reconstruction Feature Constraints

Image inpainting is the process of replenishing pixel letters within a defective area and making the restoration result visually realistic.It plays an important role in image editing,image rendering and object removal.With the proposal of Generative Adversarial Network,deep learning-based restoration methods have shown far better restoration results than traditional methods.However,the current methods still have shortcomings deficiencies when dealing with large holes.In the image inpainting task for the large area of irregular defects,it proposes an edge-guided restoration method for the problem of semantic missing and texture blurring in the restoration results;for the problem of blocked feature transfer and loss of local information during the image coding and decoding process,the attention transfer network cross-layer is proposed to improve contextual continuity and combined with U-Net for image inpainting;to improve the restoration results of structural errors and enhance the ability of characterize multi-level information in defective images,the image inpainting method based on structural constraints and multi-scale feature fusion is proposed to achieve joint improvement in structure and details.Overall,the main work of this paper is as follows.(1)It proposes an edge-guided second-order image inpainting model to accomplish image inpainting,which contains a self-encoder-based edge generation network and a U-Net-based content filling network in this paper.First,accoding to the known pixel gray values and edge information,the edge generation network infers the missing edges to obtain the edge generation map.Then,the content filling network takes the edge generation map as an a priori guide condition for texture filling.In order to reduce the loss of contextual information during texture reconstruction,the U-Net functions as a content generator,in which skip connections are utilized to combine the coded features of each level with the corresponding decoding layers.Meanwhile,U-Net introduces Shift Connection Layer to reconstruct the specified decoding layer features.To helps the content generator optimize local details,the reconstructed decoding features are fused with potential and encoding features of the same size to constrain the subsequent decoding work.(2)It proposes a generative image inpainting model based on attentional cross-layer transfer network,which applies attentional mechanism across layers in adjacent feature space to achieve fine-grained restoration effect.The model included two parts,the generator and the discriminator.First,the generator reconstructs the coding features of each layer based on U-Net in combination with Attention Transfer Network Across Layer to reduce the loss of contextual information.Then,the multiscale reconstructed features are fused with corresponding latent features for decoding using skip connections,which are applied to alleviate information loss between encoding and decoding layers thereby motivating the multi-scale decoder to generate contextually similar content.Finally,after decoding is completed,the multi-scale decoded features are converted to RGB images with L1 parametric constraints to further improve the context consistency.(3)It proposes a second-order generative image repair model based on structural constraints and multi-scale feature fusion for challenging large irregular defect repair tasks.The model is divided into two parts: an edge repair network and an image repair network.Firstly,the edge restoration network combines the self-encoder with the Dilated Residual Feature Pyramid Fusion Block to enhance the ability to characterize multi-level semantic information and structural details of the image,thus achieving higher quality edge restoration in the missing regions.Secondly,the image repair network takes real edges as a priori and embeds Dilated Multi-scale Attention Fusion Block in the self-encoder for texture synthesis to achieve fine-grained restoration of edge constraints by aggregating distant features of different dimensions.Finally,the model is trained to fuse the two networks through replacing the real edges with the edge repair results to achieve end-to-end repair from defective images to complete images.