Research Of Image Inpainting Based On Enhancement Of Edge Feature And Attention Mechanism

Image inpainting originated in the European Renaissance,with the development of digital technology and the growing needs of people for a better life,image inpainting is of great significance in people’s life,entertainment and cultural protection.At present,the methods of image inpainting work mainly from three perspectives of images: structure,texture and semantics.The traditional methods of image inpainting mainly restore the structure and texture of the damaged image,which ignores the semantics of the image itself.The current methods of image inpainting that based on deep learning could obtain high-quality images,there are still have some problems in the continuity of structure and context information of the restored images which are not highly similar to the groundtruth image.Therefore,the thesis proposes two methods of image inpainting that based on deep learning which resolves the problems of structural discontinuity and unreasonable semantics.The detailed research of the thesis is as follows:Firstly,the thesis proposes an inpainting method based on the enhancement of edge features to solve the problem of the discontinuous structure of the restored images.The deep model of image inpainting generates the edge features of the damaged areas in each loop training firstly,and uses the edge features as the priori information to guide the inpainting of the image,which guides the inpainting model to pay attention to the reconstructed edge features during learning to make up for the lost structural details,thereby improving the structural continuity of the repaired image.Secondly,the inpainting method based on the enhancement of edge feature has the problem of semantic ambiguity in the partially restored image,and the thesis proposes an inpainting method based on the enhancement of edge feature joint attention mechanism.By calculating the attention scores between the damaged areas and the known areas,which is regarded as the contribution of the known areas to reconstruct the damaged areas.At the same time,in consideration of the recurrent structure of the model of image inpainting,the thesis adds the Knowledge Consistent Attention which is an adaptive method to optimize the inpainting model.Finally,the thesis builds a human-computer interaction of web platform based on the Python language and Flask framework for image inpainting,and deploys the deep model of image inpainting proposed in this thesis to achieve the work of face editing and removal of the target object.