Non-destructive method for wood identification using conventional X-ray computed tomography data

Abstract

We establish an efficient and reliable method of wood identification that combines a non-destructive and non-invasive laboratory-scale tool, X-ray computed tomography (CT), with machine learning for image recognition. We selected six hardwood species used to create the Tripitaka Koreana and obtained the X-ray CT data of its woodblocks. Image recognition systems using the gray-level co-occurrence matrix (GLCM) or local binary patterns (LBP) were applied to the CT images and the prediction accuracies were evaluated. Because the gray level of the CT data is linearly related with the density, the CT images were preprocessed to calibrate the density. Although the resolution of the images is too low for the anatomical microstructures required for wood identification to be easily recognized visually, the predicted accuracies are quite high in both systems. However, the LBP system has slight advantages over the GLCM system. The results moreover show that the calibration of gray level to density improves the accuracies of the results. If the candidates for the wood species are selected properly and sufficient data for training is available, this technique will provide novel information about the properties of wooden historical objects.