High-precision page information extraction from 3D scanned booklets using physics-informed neural network

X-ray-based computerized tomography scans are used to analyze page information in closed booklets noninvasively. An important task is to extract the page information. Previously, the Laplace equation was used to calculate the page number field and extract the page information as an iso-surface. However, this technique cannot extract the page information properly. To solve this problem and improve the accuracy of the extracted page information, we propose a page information extraction method using a physics-informed neural network. The proposed method employs a structural similarity measure—often used in image processing research—to numerically evaluate the appropriateness of the page extraction. New history booklet is used to verify the effectiveness of this method in addition to the conventional booklet data. Graphical abstract