Vision Transformers with Contrastive Learning for Robust Wood Species Identification

Automatic wood species recognition has become increasingly primitive with the growing importance of sustainable forest management and the need for accurate identification in the timber industry. Conventional ML techniques such as Local Binary Patterns (LBP) and Gray-Level Co-occurrence Matrix (GLCM) are built-based feature extraction techniques, which have shown encouraging results. However, these approaches are often limited by factors including sensitivity to image rotation, requiring extensive preprocessing, and limited scalability across large or heterogeneous datasets. The paper proposes an innovative method to identify wood species through Vision Transformers (ViT) together with self-supervised contrastive learning techniques. The image classification methods Local Binary Patterns (LBP) and Gray-Level Co-occurrence Matrix (GLCM) need extensive preprocessing, along with being sensitive to image variations. SimCLR is employed to pretrain Vision Transformer (ViT) models on unlabeled wood images, followed by supervised fine-tuning to achieve effective classification performance. Automatic feature extraction becomes unnecessary because this method delivers consistent results no matter what the configuration. The proposed methodology achieves 99.4% accuracy on wood species classification using experimental results on datasets that contain six distinct wood species and surpasses conventional methods in performance. The system presents generalization capabilities that qualify it for practical use in forestry industries, along with manufacturing and customs inspection tasks.