Unsupervised Transformer Framework for Real-World Furniture Surface Defect Detection

In industrial manufacturing, deep learning methods have been widely adopted for surface defect detection to enhance product quality assurance. However, it remains challenging due to the high visual similarity between defects and backgrounds, as well as variations in defect scale and shape. In the present study, we introduce FUTDnet, an unsupervised transformer-based detection network, integrating two innovative designs to attain outstanding performance. First, an adaptive group-wise feature fusion strategy is proposed to aggregate patch tokens across multiple layers, which enhances semantic diversity and facilitates the accurate identification of subtle defects that closely resemble background textures. Second, a deformable cross-attention module is devised to adaptively sample informative regions while suppressing redundant background. By dynamically focusing on defect-relevant areas regardless of scale and shape variations, the accuracy of the localization is significantly improved. To enable rigorous real-world evaluation, we introduce FurniBoard, a dedicated dataset collected from furniture panel production lines, specifically targeting complex-textured surfaces with defects of varying scales-scenarios that are insufficiently represented in existing datasets. Experimental results on the FurniBoard dataset verify that FUTDnet delivers outstanding accuracy in both defect detection and localization under challenging industrial scenarios.