Research On GAN-based Ancient Text Image Restoration Network

Ancient textbooks are important reference materials for historical and cultural research and have great historical value,but their integrity and readability have been affected by long circulation and improper conservation measures,such as natural aging,mold and moisture,weather erosion,and human damage.Early restoration of ancient texts was mainly manual,which was inefficient and prone to secondary damage.With the development of the information age,the digital conversion of ancient texts into images and the restoration of defective images by computer has become an efficient and safe solution,which not only prevents the original materials from being damaged again,but also makes full use of image processing technology to improve the content of the damaged areas.At present,deep learning-based image restoration has made great progress,but the restoration of ancient text images is still full of challenges.First,there is no complete and publicly available dataset of ancient text images and Mask dataset that fits the real defects of text images.Second,the existing image restoration methods mainly target natural images,and they do not take into account the characteristics of ancient text images,which can lead to unreasonable content structure and texture blurring when used in ancient text images.At the same time,the current image restoration network restores purely at the image level,which has the problem of missing the semantic rationality of the text,resulting in seemingly reasonable but ambiguous restoration results.Therefore,based on the existing image restoration algorithms and related researches,this paper has been further explored to obtain better restoration results,and the main work of the paper is as follows:(1)In this paper,the ancient text image dataset and text image dataset are constructed for the restoration study of ancient text images.Considering the feasibility of ancient textbook image restoration,a Mask dataset adapted to the ancient textbook image dataset is also constructed to make the training closer to the real defects for image restoration and achieve the purpose of restoring real defective ancient textbook images.(2)To address the problems of unreasonable structural content and texture blurring of existing algorithms in the restoration of ancient textbook images,this paper proposes an image restoration network guided by edge features to ensure the structural rationality of the restoration results by using edge features,and introduces a multi-scale fusion block based on dilated convolution to alleviate the global and local features of images in the restoration process by using different perceptual fields to learn By using different perceptual fields to learn the global and local features of the image,the texture blurring problem occurs and the sharpness of the defective area of the image is improved.The experimental results show that the proposed algorithm can effectively recover the content of the defective region.(3)To address the problems of unclear texture and lack of semantics in ancient text image restoration,an image restoration network based on Laplace pyramid decomposition is proposed in this paper.The use of dual cross-encoders and image pyramid decomposition strategies motivate the network to generate restoration results with clear textures and alleviate the dependence of the network on edge maps to avoid the impact of edge restoration errors on the network.In addition,the introduced text learning network provides a basis for the semantic rationality of text restoration,prompting the network to generate visually complete and semantically reasonable restoration results.Experimental results show that the proposed algorithm can obtain clearer and more reasonable restoration results than existing classical restoration algorithms.