Cross-Sectional Analysis of Timber Boards Using Convolutional Long Short-Term Memory Neural Networks

Abstract

This paper proposes a one-dimensional convolutional long short-term memory (1D-CNN-LSTM) model for estimating the pith position and average ring width in Norway spruce timber boards. The model predicts these cross-sectional parameters by processing sequences of light-intensity signals derived from optical scans of the board’s four surfaces. The dataset used for training the model consists of synthetic boards sawn from simulated 3D logs. The model was evaluated on a dataset consisting of 552 end cross-sections from actual Norway spruce boards. Comparisons between the automatic and manual pith and ring width estimations demonstrated a very good accuracy. The computational speed of the model was more than twice as fast as the quickest method available in the literature. A large set of boards was then used to determine the advantages of incorporating the automatically determined average ring width in formulating indicating properties for machine strength grading. This evaluation revealed that the average ring width could, in certain situations, compensate for unknown variables such as density or resonance frequency in predicting the tensile and bending strength of Norway spruce boards.

Keywords: Sawn timber, Pith detection, Wood scanning, LSTM, Machine strength grading