Predicting Inner Wood Defects from Outer Bark Features – WoodSeer

Using supervised machine learning techniques, this approach aims to extract relevant features from 3D scans of tree bark obtained using various sensors (laser, cameras, etc.). These features are then linked with the extent of any associated internal defect, determined through densitomettry by X-ray computed tomography. The ultimate goal is to reconstruct the internal defect solely from external information.
A significant challenge of this project is the need for data covering the entire potential application domain, from standing trees in forests to the early stages of log processing in industrial settings. This involves linking all acquisitions together, as well as with X-ray-derived data describing internal defects, to enable analysis in a common 3D reference frame and provide ground truth data for training and testing algorithms.
A crucial step in the implementation of this original protocol, coupled with the development of a suitable database architecture, was the construction of a geometric scanner for logs, designed to provide a 3D colorized representation of their surfaces. The purpose is not only to provide data comparable to measurements taken in industrial environments, such as log yards, but also to enable a phase of the protocol that links measurements already taken in the forest and future measurements using X-ray tomography. After an initial description of the log's envelope, this process involves inserting small PVC cylinders radially into the log, protruding a few centimeters. From a second 3D scan, the distal ends of the cylinder axes are estimated through post-processing using a singularity detection algorithm applied to the bark surface. In the volumetric data obtained from the X-ray, the high density of PVC compared to wood allows for the segmentation of the cylinders and the estimation of the same reference points, thus defining the geometric transformation that links the two measurement reference frames.

The mobilization of 3D data and X-ray tomography from previous projects provided an initial foundation for the development of algorithms from the outset of the project. The use of deep learning has improved the performance of bark singularity detection, which may be linked to an internal defect, by 10%. AI is also effective in the automated processing of X-ray scans to isolate structural elements such as knots, particularly in softwoods. For hardwoods, knot segmentation is more challenging due to lower density contrast.
WoodSeer provides a proof of concept for new methods of assessing the internal quality of roundwood by virtually reconstructing knots from a 3D external description of the bark, using a recurrent neural network with long short-term memory (LSTM) allowing for the propagation of contours from slice to slice. These results are very promising for softwoods but still need to be improved for hardwoods.
Indeed, one of the challenges encountered is the more difficult segmentation of knots in hardwoods due to lower density contrast and more complex knot shapes, which would require future investment in labeling knots in X-ray data to train neural networks.
A database containing more than 9000 bark singularities of hardwood and softwood species is still under development, combining different 3D surface descriptions, including ground truth, and X-ray scans to establish the ultimate truth regarding the presence of internal defects.
In addition to descriptions made in the forest, the creation of the database relies on 3D measurements approaching those taken in an industrial log yard context. Despite being developed later, the geometric scanner created as part of the project, in addition to describing the log surface, allows for the superposition of all measurements taken from the forest to the X-ray scanner in the measurement reference frame of the tree in the forest, thanks to a methodology developed within the project.

The outlook is to consolidate the proof of concept with the data acquired within the WoodSeer project, which more closely corresponds to the context of a primary processing supply, while benefiting from their richness to improve algorithms: for example, the available colour information has not yet been introduced into our work.
Regarding data acquisition and the large amount of data required by AI, a path specifically reserved for the industrial context is also possible : reconstructing the internal structure of defects from information acquired on the faces of products, combined with 3D data from a true shape scanner at the beginning of the processing chain.
The perspective of establishing a link with other works on log quality or traceability carried out in previous projects (ANR TreeTrace, ADEME Biomtrace...) that focus on the ends of logs is also a necessity to propose a global approach to the quality of supply.
Other barriers must also be overcome regarding the acquisition of 3D data in the forest, the operational nature of which largely reserves it for R&D activities. Nevertheless, technologies are evolving very rapidly, and our role would be to establish the specifications linking acquisition quality and estimation quality. In the same spirit, the necessary computing resources must also be considered to envisage portable solutions, and both acquisition and processing must be optimized.
Despite the work that remains to be done, the results obtained on softwoods are beginning to open up the possibility of setting up a demonstrator in an industrial environment.
In terms of funding to continue, the PEPR Forestt call for proposals is a current target, following the submission of a declaration of interest deemed eligible.

Khazem, S.; et al. Improving Knot Prediction in Wood Logs with Longitudinal Feature Propagation. LNCS. 2023, 14253, 169-180.

Khazem, S.; et al. Deep learning for the detection of semantic features in tree X-ray CT scans. Artif. Intell. Agric. 2023, 7, 13-26.

Delconte, F.; et al. CNN-based Method for Segmenting Tree Bark Surface Singularities. IPOL 12, 2022, 1-26.

Delconte, F.; et al. Tree Defect Segmentation Using Geometric Features and CNN. In: RRPR 2021, LNCS, LNIP 12636, 80-100.

WoodSeer is a project aiming the development of automated means for characterizing wood supply in the upstream part of the forest-wood chain in the context of inventory, log trade and first processing, by providing detailed information on the nature, location and dimensions of internal defects in an automated manner. Nowadays, for this industry, this information is only accessible through the use of X-ray CT scanners providing an internal reference structure for the rameal system, but not easily affordable for companies because of the high required investments. Nevertheless, this information is integrated through the experience and know-how of experts in the wood grading process in forests or wood-yards. Thus, the founding idea of the project is to link the signatures of defects on the bark with their internal characteristics. Without being exclusive, the targeted defects are the knots of the pruned branches included in the trunk. Indeed, the scar of a branch on the bark contains information about the height of the bud that gave birth to it, the diameter of the branch at the time of pruning, its average inclination and the radial growth of the trunk after pruning.
Thanks to previous works involving several project partners, very promising initial results were obtained on the detection of defect signatures on the surface of smooth or rough bark, using algorithms processing the 3D description of the log surface made with a terrestrial Lidar. This work also provided a classification of the defect type and the computation of relevant characteristics.
In continuity and in a first Work-Package, the WoodSeer project aims to study a diversification of the digitizing methods able to provide a three-dimensional description of the surface of a log through the use of portable devices in the forest and on industrial sites using true shape scanners, but also to assess their relevance in detecting surface defects. Some tasks are dedicated to provide data to link the external and internal description of the same defect, through the use of an X-ray scanner. Others tasks aim to complete the datasets with synthetic 3D geometric structures in order to be able to use efficient artificial intelligence methods, such as Deep Learning in the other tasks of the project.
The second Work-Package aims at improving the available methods for detecting and classifying surface defects by exploring other ways, such as combining Deep Learning and geometric constraints to facilitate the reinforcement of relevant databases in the future.
The last Work-Package is dedicated to the establishment of external-internal relationships, in order to be able to predict the internal defect from its external signature, with the ambition of relying on artificial intelligence methods such as neural networks.
To meet these objectives, the consortium gathered in this project benefits from a representativeness of the sector thanks to managers and sales representatives of the public and private forests, a sawmill, a subcontract with a company designing and assembling geometrical shape scanners and buckling optimization software. On the scientific level, WoodSeer brings together recognized expertise in 3D surface and volume digitization and their processing to extract information. The project also includes skills in tree development, its impact on wood quality, and in statistical modelling, as well as skills in discrete geometry, synthesized objects, and finally in artificial intelligence.