Wood Classification using Convolutional Neural Networks and Vessel Features on Limited Dataset

Abstract

Wood identification can be considered as one of the essential tasks for experts on wood anatomy. The task of wood identification is not only required by wood anatomists, but it is also needed in other fields, such as custom ports, forest survey, and wood industries. However, the availability of wood identification is limited due to the number of wood anatomists. A computer vision-based algorithm is proposed to perform wood classification to resolve the existing problems. The algorithm uses an image of wood surfaces to determine the wood types. A trained identification model is applied to classify the wood name based on a convolutional neural network (ConvNet) algorithm. The images were acquired using a smartphone camera and additional loupe mounted on the camera lens. The paper focuses on twelve (12) wood types that have a visual similarity appearance. Two classification models based on a convolutional neural network have been developed and used to obtain an acceptable accuracy on the observed wood types. The first model, Model-1, is created by stacking the convolutional and flattening layers. Meanwhile, the second model, Model-2, combines the unsupervised clustering technique and the ConvNet algorithm. The results show that both models can perform well, indicated by an F1 score ≥ 0.90, at five species. The F1 scores of the three species are still found to be close to 0.50 at three species. Specific techniques need to be explored and applied for these species. The methods might improve the classification performance.