SUPR
Procedural generation of wood using AI
Dnr:

NAISS 2024/5-195

Type:

NAISS Medium Compute

Principal Investigator:

Osama Abdeljaber

Affiliation:

Linnéuniversitetet

Start Date:

2024-09-01

End Date:

2025-09-01

Primary Classification:

20103: Building Technologies

Webpage:

Allocation

Abstract

Nondestructive evaluation of timber properties such as strength and stiffness is vital for the sawmilling industry. Accurate strength grading of timber boards leads to more effective material utilization. The main factor that influences the mechanical performance of timber is the presence and shape of knots, which cause deviations in the wood fibers. Moreover, in softwood species, the distance from the center of the tree (pith) affects various mechanical and physical properties of the wood. Therefore, to model and grade sawn timber accurately, it is necessary to know the knot geometries and pith locations. Deep learning (DL) offers great opportunities for developing automatic systems for the identification of knots and pith locations by processing high-resolution images obtained using optical or X-ray scanning of timber boards. Nevertheless, training supervised DL models requires obtaining large amounts of manually labelled training data from scanned boards, which is both costly and time-consuming. One way to overcome the difficulties of obtaining labelled data from actual timber boards is to generate synthetic boards with realistic surfaces and then use them to train the DL models. We have previously demonstrated that virtual boards generated by a rather simple model that only generates clear wood (i.e. knot-free) timber are sufficient for training an accurate DL model for pith detection. In this project, we aim to develop an advanced generative AI that can produce random 3D wood logs with realistic features such as knots, annual rings, and fiber orientations. We will use an unlabeled dataset consisting of hundreds of 3D CT images from X-ray scanning of full-scale timber logs to train this model. We will experiment with different types of generative models, such as Generative Adversarial Networks (GANs) and Diffusion Models. We expect that the 3D boards generated by our AI will enable us to train various DL models for predicting knot geometries, pith locations, and fiber orientations, leading to more precise strength grading and better utilization of timber.