2024 Volume 70 Issue 3 Pages 63-71
Measuring the drying checks requires a lot of time and effort, and the judgment of the drying checks differs depending on the person. Semi-automatic measurement methods based on image processing in previous studies required special treatment of specimens, and has not solved the problem of time-consuming. In this study, to develop a convenient measurement method to quantify drying checks in sawn timber, we created a Convolutional Neural Network (CNN) that can discriminate drying checks in cross-sectional images pixel by pixel. The CNN was trained on 850 cross-sectional images (cross-sectional dimension 105 × 105 mm) of sugi boxed-heart square timber acquired by a flatbed scanner, and a testing set of 150 images evaluated its performance. The results showed that the Root Mean Squared Error (RMSE) for total check area and total check length were 11.8 mm2 and 28.4 mm, respectively. In especially, the area was found to be predictable with high accuracy. The low accuracy of length prediction compared to area was attributed to the loss of small checks information when the images were compressed and input into the CNN. It was necessary to change the method of inputting images to the CNN to improve the prediction accuracy.
