earticle

영어

In order to improve the accessibility of wood species identification, the four domestic and six imported softwood species were classified using deep learning method. The cross-section micrographs were used as a dataset; which 1,535 images of 40x micrographs with earlywood and latewood, and 2,000 images of 200x micrographs for earlywood and latewood each. The classification accuracy and loss rate of 10 species were compared using the four convolutional neural network models such as modified CNN, GoogLeNet, VGG16, and ResNet. In verifying the classification accuracy and loss rate by model, the influencing factors, such as epochs, collected part of the dataset, and dataset augmentation, were analyzed. The modified CNN and GoogLeNet models increased classification accuracy in proportion to the number of epochs, achieving more than 95% classification accuracy in the final stage. At the same time, the loss rate decreased with decreasing the number of epochs. VGG16 model showed a low classification accuracy of 20~30% and a high loss rate regardless of the number of epochs during learning under the same conditions as the other models. The ResNet model showed a high classification accuracy of over 90%, with a low loss rate during the training process. However, classification accuracy decreased to 20~30%, with a high loss rate when the test process. As a result of analyzing the general trends in the four models, the classification accuracy increased with increasing the number of epochs in the latewood and total dataset. In contrast, the earlywood dataset didn't show any tendency. The dataset augmentation was not significantly correlated with the classification accuracy and loss rate. Based on these results, modified CNN and GoogLeNet models among the four deep learning models showed excellent wood species classification performance. Further, it is expected that the two models can be applied to identify unknown ten softwood species.