Application of deep convolutional neural network on feature extraction and detection of wood defects

Abstract The artificial extraction of features from wood images via a conventional method is complex. Therefore, we proposed a learning method to detect wood features and automatically classify defects from wood images that were collected using a laser scanner through a deep convolutional neural network (DCNN). We applied TensorFlow to train the network, which was composed of an input layer, four convolutional layers, four max-pooling layers, three full-connected layers, a softmax layer, and an output layer. We used dropout, L2 regularization, and data augmentation to avoid the overfitting problem. For the training and evaluation of the DCNN model, a wood defect dataset was collected from 600 pieces of red and camphor pine wood and then divided into training (including 42,750 knots, 40,050 cracks, and 41,200 mildew stains), validation (including 6000 knots, 6000 cracks, and 6000 mildew stains), and testing (including 6000 knots, 6000 cracks, and 6000 mildew stains) datasets after data augmentation. Finally, by using this improved DCNN model, we achieved an overall accuracy of 99.13%; that is, only 1.12 s was needed for detection (including image preprocessing and identification). These results showed that the proposed DCNN model could recognize wood defects more accurately and effectively than conventional methods when using wood images.