Classification of scots pine (Pinus sylvestris) knots in density images from CT scanned logs

Abstract

This study was carried out to test methods for separating knots from clearwood in a digital image stack when scanning for internal defects with a medical CT-scanner. Scots pine knots, represented by its tangential surface density image extracted from a CT-image stack, have been classified by two different methods showing equal results. The knots are classified in four knot types by an Artificial Back-propagation Neural Network (ANN) and a Partial Least Squares Modelling with Latent Variables (PLS) model. The classification precision of aim of four different knot types, is between 85% and 97%. The results indicate that both methods may be useful tools in order to describe and classify knots in concentric surfaces around the pith in CT-images and thereby extract parametrical models from CT raw data image stacks. A simplified classification model has been obtained, by analysing the learning patterns in both the ANN and PLS model, that classify knots and transform density related data to segmented and classified parametrical descriptions.