Deep learning for use in lumber classification tasks

Abstract

Digital image processing has been widely used in the wood industry, and it is expected that the algorithms are accurate and have low computational complexities, especially for the real-time lumber grading and lumber classification processes. This paper investigates variations of deep learning strategies based on ResNet18 for classification of lumber images. The four datasets used in this work were manually marked as lumber defects, wood textures and lumbers by experts. A key ideal is to employ the transfer learning in the context of convolutional neural networks with a classifier layer only training with a small amount of training data for different tasks at the same lumber machinery. Through the expansion of unbalanced samples, the accuracy rate has been effectively improved. The human involvement when needed is kept to a minimum only for the training phase. The proposed approach was independently tested with four datasets, of which 80% of the data is used for training and 20% of the data is used for testing. The classification accuracy of the approach for each of the datasets is 98.16%, 93.32%, 96.64% and 99.50%. The average time for sorting the lumber image was kept at 0.003 s when the system runs on Nvidia GTX860 GPU.