Open Collections

Mason, Lewis

University of British Columbia

The pulp and paper industry explores fibre mixtures to improve products, reduce costs, conserve energy, and access new markets. X-ray micro-computed tomography enables the micron-scale visualization of paper handsheets, allowing per-fibre analysis at unprecedented detail. Iron coated cellulose fibres, which closely resemble regular fibres but appear slightly brighter in tomography scans, have been developed and mixed into handsheets, allowing a new form of per-phase analysis that was not possible before. This poses the challenge of distinguishing tracer fibres from normal fibres—known as the ”tracer problem”—which is difficult for traditional software algorithms due to the similarity of both phases in shape and intensity. To address this problem, a custom convolutional neural network architecture with a novel layer, K-Origins, is designed using two dimensional (2D) synthetic data that resembles tomography scans. The K-Origins layer is demonstrated to drastically increases the segmentation accuracy for two cases: Object detection and tracer segmentation. A custom pipeline is then implemented with these findings and is used for cellulose tracer fibre segmentation on a pre-existing dataset of two mixed handsheet scans; 0.1 wt% and 2.0 wt% tracer fibre to normal fibre ratio handsheets. The segmentation accuracy for these scans is improved by 7.45 and 2.82 times, respectively, compared to the previous best technique which requires oil-submersion prior to scanning. Additionally, the custom layer K-Origins likely has applications in many other fields that use convolutional neural networks.

Text

2025-01-13

Attribution-NonCommercial-NoDerivatives 4.0 International

http://hdl.handle.net/2429/90112

Mechanical Engineering

University of British Columbia

UBCV

http://creativecommons.org/licenses/by-nc-nd/4.0/