| Title |
3-D Movable Type Reconstruction from Old Printed Documents using Deep Learning-based Character Extraction and Recognition |
| Authors |
정범채(Beom-Chae Jeong) ; 최강선(Kang-Sun Choi) |
| DOI |
https://doi.org/10.5573/ieie.2022.59.9.74 |
| Keywords |
Type reconstruction; Image segmentation; Character segmentation; Character classification; Historical printed documents |
| Abstract |
Hangul movable types and historical printed documents have important values in various aspects, such as printing history, Korean history, and Korean font, as well as historical values. However, most of the movable types at that time have not been fully preserved and handed down. In this paper, we propose a method of reconstructing the 3D type by extracting type structure information from the printed documents using deep learning. First, U-Net is applied to scanned high-resolution image data of printed documents to extract character stroke areas. Each stroke region is classified to a phoneme unit to recognize the character. The extracted character segmentation data classifies the same characters through a convolutional neural network (CNN) model, which is recognized by dividing them into phonemes. After matching between the same character images using a homography network simplified to obtain movement and rotation transformation, it is subdivided into characters printed in the same type based on similarity. Finally, each 3-D movable type model is reconstructed by integrating the character images corresponding to an identical movable type. Through the experiment, it was confirmed that each proposed process successfully performs. |