On reflections of wood : wood quality features modelled by means of multivariate image projections to latent structures in multispectral images

Sammanfattning: Classifying logs and boards has an essential economic impact on the forest products industry. In a modern production chain, quality control and raw material specifications are essential to fulfill production system demands as well as end user demands. The subject of the thesis is automatic grading of wood represented by its appearance in various stages of decomposition of a tree, i.e. as logs, boards, components or individual surface objects often labelled as defects. The main method used is based on multivariate measurements and data compression by means of principal component analysis (PCA) and prediction modelling with projections to latent structures (PLS). The thesis adresses mainly feature identification and feature extraction of properties that have an impact on automatic grading of sawn softwood boards. The features are modelled by means of their appearance in the multivariate image space obtained by scanning and integrating images utilizing different imaging sensors. The main goals of this thesis are: to introduce multivariate image analysis (MIA) and multivariate image projections to latent structures (MIPLS) combined with experimental design as a research tool for wood feature description and to give a survey of its applications to identify and extract features feasible for classification of softwood surfaces to investigate the use of an imaging spectrograph combined with a smart sensor for real time grading of sawn products to present the necessity of installing human evaluation and system interaction possibilities into automatic wood inspection systems to obtain good performance This thesis contributes to the wood research field by the multivariate approach so far with a limited use in wood technology research with a few outlayers. The main contributions are: the introduction of PLS and MIPLS as soft modelling tools for calibration and prediction modelling of wood features the evaluation of the imaging spectrograph/MIPLS concept to model the spectral behaviour of the interaction between visible light, the wood surface and the sensor the integration of a multisensor, i.e. a combination of sensors such as X-rays, R-waves and CCDs, and how to insert the obtained data into a linear or non linear model utilizing sufficient information from each sensor the recognition of useful information in softwood scanning. Utilizing electromagnetic wave interaction based sensor data, revealing the latent variables and their variations with respect to quality grading.