Optimizing the Parameters of a Knot Assessment Model Based on the Visual Grading of JAS of Lumber

Abstract

We defined a knot assessment model based on the visual grading of Japanese Agricultural Standard (JAS) of lumber and attempted to optimize parameters of the model using simulated annealing. Parameters of the knot assessment model were optimized by maximizing coefficients of determination of regression models, using data of bending tests and surface images of 40 test pieces, 120 mm square cross section kiln-dried lumber with pith of Japanese larch (Larix kaempferi Lamb.). Values of the coefficient of determination of regression models using knot index values of optimized knot assessment models were much larger than those obtained by ordinary visual grading. However, knot index values of almost all test pieces were decided by a particular type of knot diameter ratio, suggesting overfitting. We then simplified the knot assessment model, eliminating the distinction between center or edge areas and types of single knot. Values of the coefficient of determination of regression models were then decreased compared to those of the above optimized knot assessment models. However, they were still larger than those based on ordinary visual grading, and there were significant differences between correlation coefficients of each single regression model.