Wood Artificial Intelligence — Knot modeling by Computed Tomography

Wood can serve as long-term carbon sink, and increased use in buildings is a means to reduce the negative effects of climate change. Consequently, it is important to produce sawn timber suitable for buildings, furniture and other products with long lifespans. As more wood is used, it is even more important to use the wood resource efficiently. Industrial computer tomography (CT) scanning of logs is becoming a more widespread tool in sawmills, and it has proven to considerably increase the volume of high-value sawn timber. CT scanning of logs provides information on the interior properties prior to sawing, and this can be considered when optimising the sawing. The most difficult CT-detection problems are the size of knots in sapwood and the transition between sound and dead knot sections. Knot size and whether knots are sound or dead are important factors for sawn timber quality, and the requirements vary between different products. Knot size is a limitation for structural timber since it has negative effects on strength and stiffness. For timber used in visual applications, where the appearance is important, there are additional requirements for sound and dead knots. A more accurate detection of knot properties in CT scanning will improve the quality prediction and contribute to a more efficient use of the wood resource. The knot properties are related to the crown development and branching of the trees and can be predicted from external tree measurements and forest inventory data. The project aims at improving the algorithms for knot detection in CT scanning by use of artificial intelligence (AI), combining data from CT scanning with physical measurements of knots and forest inventory data, and developing new models for estimating knot properties and strength and stiffness of sawn timber. Materials of 24 Scots pine trees and 24 Norway spruce trees are collected from 16 locations in three regions in Sweden. The logs have been CT-scanned in wet condition, and samples of branch whorls have been extracted and scanned again after drying to about 30% moisture when the knots in the sapwood are easier to detect. In total, 1843 knots have been recorded in Scots pine and 1751 in Norway spruce. Geometrical measures of the knots have been recorded, and now we have a complete dataset with forest- and tree data, CT-data, and data from physical measurements of the knots. A corresponding material of Norway spruce has been sampled in Germany. As mentioned earlier, we have two stages of CT scans: wet and dry versions of the same disks. The dry scans provide better contrast between the knots and the surrounding wood, and our aim is to begin segmenting the knots from the dry data. However, in industry situations, we need to account for the wet data and perform segmentation on that. To achieve this, we require registration software to match both datasets so that we can transfer information between them. We have developed a semi-automated Matlab-system for this purpose, which uses matching algorithms to align the data. In most cases, some manual adjustments are needed to achieve perfect matching. We have already matched the data for both Scots pine and Norway spruce. Regarding segmentation, we are in the process of creating ground truth data to start modeling the segmentation models of knots and the sound-dead knot border in the disks. Currently, we have an almost developed Python software to support this work. In this software, we can load both the wet and dry versions of the data together. We have established a setup for physical measurements over the CT data, which aids in creating the best possible ground truth during annotation. By using the registered data, we can transfer the annotations made on one version to the other. As soon as we have the ground-truth data ready, we can start creating segmentation models.

Knot-extraction methodology from CT log scans will be developed with more details and a higher accuracy than those of existing algorithms. This will correspond to a significant improvement over existing methods. Industrial CT scanning of logs is becoming a more widespread tool in sawmill production and has proven to considerably increase the volume of high-value sawn timber. Still, there is great potential in improving its accuracy. The novel image analysis methods based on deep learning and neural network approaches, in combination with improved models, will enable a more precise production optimisation from forestry to log and final end-product (sawn-timber). The project is divided into 9 WPs covering different levels of the processing chain, centering the branch/knot structure, and linking tree (crown, growth) and mechanical wood (strength, stiffness) properties. A unique database will be created and will be used for the following purposes: - AI-based algorithms for better extraction of knots from CT images, combined with statistical modelling based on the training data sets. - Combination of forest inventory data and detailed tree growth data, external stem, crown and branch measurements and knot properties with data from CT scanning to improve sawing simulations. It will also make systems for more optimal bucking of stems possible, and feedback to forestry and silviculture, and therefore the forest value chain will be more sustainable due to a higher material efficiency. - Systematically compare CT measurements of stem/knot structure in green and dry conditions with manual, physical measurements of the same properties in order to improve algorithms for knot extraction from CT images. - Development of new grading models for estimating strength and stiffness of virtual boards based on CT data.