Deep learning segmentation of wood fiber bundles in fiberboards

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