Background and Objective

To support deforestation-free supply chains, the wood species used in products are determined at the Thünen Centre of Competence on the Origin of Timber. Routinely, cut specimens of wood and, in the case of fibrous materials such as paper, isolated cells are examined under a light microscope and analyzed by experts in a time-consuming process. Automated image recognition systems can significantly reduce the workload and the time required.

In the project in cooperation with the Fraunhofer Institute ITWM, machine-learning systems are being trained. The successfully developed systems for the recognition of hardwoods from the previous project KI_Wood-ID will be supplemented by additional softwood systems in the WoodFiberID project. Together, these will be combined in an image evaluation software that can be used to analyze microscopic images of paper specimens. The results will be published scientifically and provided to all testing institutes for the control of internationally traded wood.

Target Group

Authorities, science, timber trade and consumers

Approach

For a selection of relevant timbers, samples from the scientific collection and other documented sources are processed into realistic training and test specimens. In the slide scanning microscope specially acquired for the previous project, 100 slides can be inserted simultaneously. Automated microscopic images are then taken in multiple focal planes over the entire slide area. These images represent the important data basis of the project. In these images, the characteristic morphological features that are essential for the identification of the wood genus by wood anatomists are labeled. Machine learning systems are trained and validated with these at the ITWM. 

Microscopic images of fibrous materials can already be analyzed by the first prototype user interface developed in the previous project. The machine-learning systems can already identify eleven hardwood species with whose references they have been trained. Below you can see how the cells that are important for identification are first detected in the image and then classified. The identified cells can then be checked for plausibility by wood anatomists.

In the current project, the focus is now on softwoods and non-wood raw materials such as grasses and sugar cane bagasse, which are increasingly being used in fibrous materials. For softwoods in particular, completely independent models have to be trained because a completely different cell type has to be used to identify the type of wood.

Our Research Questions

Can machine learning also identify softwoods and non-wood fibers in fibrous materials? 

Links and Downloads

Project partner Fraunhofer institute ITWM