Image Classification of Imbalanced Wood Microscope by Integrating Multi-source Features

This is the authors' abstract. We don't add key points for this paper.

Owing to factors such as human activities and environmental pollution, the global forest area is continuously diminishing. Consequently, rapid and accurate identification of wood species becomes extremely crucial. With the recent advancements in deep learning, wood classification based on digital images is currently relatively popular method. Nevertheless, there still exist some challenges, such as the difficulty in integrating plant phenological and texture features and the inability to automatically select features of inter-class images. This paper presents a multi-source feature fusion classification, which combines traditional local feature description with automatic feature extraction by deep learning. The proposed method reduces misclassification and addresses the imbalance problem in wood data. Besides, the method is validated on the microscopic cross-sectional images of 75 broad-leaved wood species, with an accuracy of 94.0%, indicating its potential for application in the classification of imbalanced wood data. The comparison experiments with other existing models on the wood dataset demonstrate the effectiveness of the proposed method.