Deep learning segmentation of wood fiber bundles in fiberboards

Natural fiber composites and fiberboards are essential components of a sustainable economy, making use of biosourced, and also recycled materials. These composites' structure is often complex, and their mechanical behavior is not yet fully understood. A major barrier in comprehending them is the ability to identify the fibers in situ, i.e. embedded in complex fibrous networks such as medium-density fiberboards (MDF). To that end, the first step is to separate individual wood fibers from fiber bundles. Modern material studies on real world, dense fibrous materials using X-ray microtomography and 3D image analysis were always limited in accuracy. However, recent machine learning techniques and particularly deep learning may help to overcome this challenge. In this work, we compare existing segmentation algorithms with the performance of convolutional neural networks (CNNs). We explain the need for network complexity, and demonstrate that our best algorithm, based on the UNet3D architecture, reaches unprecedented accuracy. Moreover, it achieves the first segmentation sufficiently qualitative to extract morphometric measurements of the fiber bundles and accurately estimate their density. Among other applications, the proposed method thus enables the design of more realistic material models of MDF, and is a milestone towards the understanding and improvement of this wood-based product.

Automated summary

Natural fiber composites and fiberboards are important components of a sustainable economy, but their complex structures and mechanical behaviors pose challenges in identifying and analyzing the fibers. In this study, deep learning algorithms were used to separate and measure individual wood fibers, achieving unprecedented accuracy and providing a milestone for designing more realistic material models and improving wood-based products.