A Study into Methods for Classifying Cellulose Fibre Distributions in Paperboard

Sammanfattning

This Master’s Thesis explored how traditional computer vision methods and deep learning methods could be used to estimate hardwood and softwood fibre distributions in paperboard. 11 mixtures ranging from 100% hardwood to 100% softwood with 10% increments were analysed. The study compared the usefulness of tomograms and radiographs for the given task. It was found that traditional methods did not identify differentiating features or reliable trends in the tomograms. The tomogram dataset only consisted of two volumes per mixture which made it difficult to generalise the results. The radiograph dataset was larger and combined with pretrained models gave the best results. As the pretrained models found differentiating features in the radiographs, the tomograms were deemed to be unnecessarily complex for the task. Pretrained models were used to extract features from the radiographs, which were then classified into 11 classes using a custom trained MLP. Three ResNet architectures were compared and a hyperparameter search was performed on the MLP. The final model used the ResNet18 and achieved 80% accuracy and 0.80 RMSE on test data. Most incorrect classifications occurred between neighbouring classes. This is likely due to label uncertainty. Each radiograph only covers a small region of the original paperboard sheet and therefore does not accurately represent the distribution of fibres of the larger sheet. The generalisation of the model could not be verified because a directionality not visible to the human eye was detected in the data. Only 90 degree rotations were applied to the images meaning a directionality bias could still remain. It would be necessary to perform additional imaging to evaluate the model.