The image correction of ancient literature documents is a key link in the digitization of ancient literature documents,which has important practical significance for improving the quality of ancient literature digitization. This paper proposes a method for correcting ancient book document images based on deep learning and 3D feature information extraction,in response to the problems of complex deformation and difficult correction caused by oxidation bending,adhesive folding and special binding methods commonly found in ancient books. Firstly,use a U-Net encoder decoder to extract the three-dimensional features of ancient document images. Then,based on the Transformer model,perform backward mapping on the obtained 3D feature map. Finally,use bilinear interpolation to obtain the corrected image. To verify the effectiveness of the proposed method,experiments were conducted on two self-made test sets. The experimental results show that this method reduces Local Distortion (LD) by 2.61％～6.58％ compared to the DewarpNet model. The experimental results show that the proposed method can effectively complete the task of correcting ancient book document images and improve the digital quality of ancient books.
ancient book imagesdocument image correction3D information extractionTransformerencoder-decoder
NETL
NSTL
万方数据
