Segmenting knots in X-ray CT scans of logs

Abstract

The main objective of this thesis is to solve the problem of segmenting wood knots within and on the surface of wood logs by using three-dimensional (3-D) image data of wood logs obtained from X-ray Computerized Tomography (CT) scans. It is imperative to note that, prior thorough information about the internal and external positions and proportions of wood knots can be very useful for sawmill businesses due to it aiding them to plan the sawing patterns before the actual sawing process occurs which could ultimately increase their profitability. The proposed method in this thesis addresses this problem by creating a machine learning model that can segment the knots by using 3-D X-ray CT scan images of wood logs as input. The particular machine learning model that was implemented is called U-Net, which a encoder-decoder type model based on Convolutional Neural Network (CNN). The predicted segmented images found as output from the model were assessed with different evaluation criteria by comparing the ground truth data that was created with human supervision. The results from the evaluation results showed that the predicted model managed to achieve acceptable levels of overlap with the ground truth.