Identification of Fibers in Micro-CT Images of Paperboard Using Deep Learning

Abstract

This master thesis project explores the possibility of using deep learning
to segment individual fibers in three-dimensional tomography images of paperboard fiber networks. We test a method which has previously been used to
segment fibers in images of glass fiber reinforced polymers. The method relies
on a neural network which produces an embedding for each voxel in the input
image, such that the embeddings corresponding to a given fiber should form
a cluster in the embedding space. Individual fibers can then be identified by
applying a clustering algorithm to the embeddings. Although the method is
able to identify some more easily distinguished fibers, the achieved accuracy is
insufficient. We find that the main difficulty lies in acquiring training data of
high enough quality, and that future work concerning this task is required. In
this work, the use of different types of data, including synthetically generated
data, and what we refer to as a type of semi-synthetic data, have been tested.
Although we do not reach any satisfying results, this work may serve as a base
for future research.

Popular Abstract

Knowledge about the mechanical behaviour of paper is important for the paperboard packaging industry. This knowledge can help with reducing material waste, and improving the durability of packages. Paper consists of wood fiber networks, and as the mechanical behaviour of the material largely depends on the properties of these networks, it is of interest to study them.

A widely used tool for analysing materials such as paper is X-ray tomography, which can provide 3D-images of the internal structure. This master thesis project has been about exploring the possibility of using artificial neural networks to identify the different individual fibers in such an image, i.e. ’segmenting’ it. Segmenting the image would mean to identify which voxels (3D pixels) belong to each fiber. This would be a useful tool for the analysis of tomography images, as it would give researches access to many properties of the fiber networks, such as fiber lengths and thicknesses.

In the last decade, artificial neural networks have become widely used in the fields of image analysis and computer vision. To be able to perform tasks such as image classification or segmentation, an artificial neural network is trained on a set of training data. In the case of segmentation, this essentially means that the neural network is shown a number of examples of how different images should be segmented. By the use of advanced algorithms, the network should learn from these examples, so that it is able to segment other images by itself.

In this project, different methods and models for segmenting images of paper fiber networks have been tested. The conclusion is that more future work is needed to reach any satisfying results. Much of the challenge lies in the difficulty of obtaining good enough training data, which is due to the fact that the fibers in the images often are very difficult to distinguish from one another, even for a human observer.

• author: Rydgård, David
• supervisor: Sara Johansson; Johan Tryding
• organization: Solid Mechanics; Department of Construction Sciences
• course: FHLM01 20222
• year: 2023
• type: H3 - Professional qualifications (4 Years - )
• subject: Technology and Engineering
• keywords: Fiber networks, Paperboard mechanics, Deep learning, Tomography, Image analysis
• publication/series: TFHF-5000
• report number: TFHF-5254
• language: English
• id: 9110236
• date added to LUP: 2023-02-14 12:32:03
• date last changed: 2023-02-14 12:32:03