Public defence in building technology: Tadios Sisay Habite

Thesis title:

Pith location and annual ring detection for modelling of knots and fibre orientation in structural timber – A deep-learning-based approach

Third-cycle subject area:

Building technology


Faculty of Technology


Friday 26 August 2022 at 09:00

External reviewer:

Associate professor Julie Cool, The University of British Columbia, Canada

Examining committee:

Professor Jan-Willem Van de Kuilen, Technical University of Munich/Delft University of Technology, Tyskland
Associate professor Phuc Ngo, Université de Lorraine, Frankrike
Adjunct professor Mikael Perstorper, Karlstad universitet


Professor Björn Johannesson, Department of building technology, Linnaeus University


Professor Anders Olsson, Department of building technology, Linnaeus University

Assistant supervisor:

Associate senior lecturer Osama Abdeljaber, Department of building technology, Linnaeus University, Associate Professor Jan Oskarsson, Linnaeus University, Professor Welf Löwe, Department for Computer Science and media technology, Linnaeus University


Professor Thomas K. Bader, Department of building technology, Linnaeus University


Monday 20 June 2022 at 09:00 at University library, Växjö

Public can join via Zoom, a link will be published here two hours before the defence.


Detection of pith, annual rings and knots in relation to timber board cross-sections is relevant for many purposes, such as for modelling of sawn timber and for real-time assessment of strength, stiffness and shape stability of wood materials. However, the methods that are available and implemented in optical scanners today do not always meet customer accuracy and/or speed requirements. The primary purpose of this doctoral dissertation was to gain an increased knowledge and a better understanding of how different characteristics and surface defects of timber boards can be identified automatically and robustly. The secondary purpose was to explore the possibilities of how such identified features/defects can be used to add value to the wood manufacturing industry.

In the present study, three different methods were developed to non-destructively and automatically detect annual rings and pith location based on images obtained by optical scanning of the four longitudinal surfaces of the timber board. In the first method, a signal-processing-based approach and an optimisation algorithm were applied. In the second method, a deep-learning-based conditional generative adversarial network (cGAN) and a shallow artificial neural network (ANN) were used. In the third method, a single step deep-learning approach with a one-dimensional convolutional neural network (1D CNN) was applied. A novel stochastic model was also proposed to generate an unlimited number of virtual timber boards, with photo-realistic surfaces and known pith location, by which the proposed 1D CNN was trained before it was successfully applied to real timber boards. Concerning accuracy, all the three methods gave prediction errors of the same magnitude, between 4 mm and 6 mm. The 1D CNN method needed only 1.1 ms to locate the pith at a single section, which was 165 and 127 times faster than the signal-processing based and the cGAN based methods, respectively. Hence, the 1D CNN method proved to be the fastest, most operationally simple and robust method.

In sawn timber, the presence of knots causes the fibres to deviate from the longitudinal direction of the board, leading to a significant reduction of strength and stiffness. In the current study, a computer algorithm was proposed to detect knots on board surfaces and to reconstruct the knots in three dimensions (3D) by using the detected pith location. Moreover, a fibre modelling method was also proposed and used to produce the 3D fibre orientation within the volume of timber boards. Furthermore, the detected pith location and annual rings visible on the board surfaces were also utilised to estimate the radial annual ring profiles along the longitudinal direction of timber boards.

Keywords: Sawn timber, Pith location, Deep learning, Artificial neural networks, Convolutional neural network, Conditional generative adversarial network, Knot detection, Knot modelling, Knot reconstruction, Fibre orientation, Annual ring profile