Efficient Adaptation of Visual Foundation Models for Wood Defect Segmentation via Instance Linking and Feature Disentanglement

Automated visual inspection (AVI) for wood defect detection plays a crucial role in ensuring the quality of wood products in industrial manufacturing. The interclass similarity and fuzzy boundaries of defects, which do not fit neatly into predefined categories, cause current specialized deep learning segmentation methods to perform unsatisfactorily. To address these challenges, this study leverages the robustness of large-scale visual foundation models (VFMs) and proposes a layer-wise adapter module (LAM) to adapt their feature extraction backbones for high-quality wood defect segmentation. Our LAM integrates the instance-linking token module (ILTM) for enhancing sensitivity to fuzzy boundaries. Additionally, the feature disentanglement module (FDM) is designed for reducing feature redundancy, while the layer switch module (LSM) facilitates dynamic feature optimization by connecting different layers. Extensive experiments on the rubber wood and pine wood datasets demonstrate that our LAM significantly improves both segmentation accuracy and efficiency compared to ten state-of-the-art methods. Cross-dataset evaluations further validate its robustness in generalizing to unseen datasets, even under substantial domain shifts. Additionally, qualitative analyses on the OULU-DET and VSB-DET datasets have proved LAM’s zero-shot transferring ability to effectively segment unseen defects without the need for pixel annotations.